HAL Id: tel-01558141 https://tel.archives-ouvertes.fr/tel-01558141 Submitted on 7 Jul 2017 HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers. L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés. Atomization and dispersion of a liquid jet : numerical and experimental approaches Francisco Felis-Carrasco To cite this version: Francisco Felis-Carrasco. Atomization and dispersion of a liquid jet : numerical and experimental approaches. Fluids mechanics [physics.class-ph]. Ecole Centrale Marseille, 2017. English. NNT : 2017ECDM0001. tel-01558141
152
Embed
Atomization and dispersion of a liquid jet: numerical and ...Xavier DEMOULIN for the initial OpenFOAM support at CORIA, Pr. Jean-Philippe MATAS for the experimental post-treatment
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
HAL Id: tel-01558141https://tel.archives-ouvertes.fr/tel-01558141
Submitted on 7 Jul 2017
HAL is a multi-disciplinary open accessarchive for the deposit and dissemination of sci-entific research documents, whether they are pub-lished or not. The documents may come fromteaching and research institutions in France orabroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, estdestinée au dépôt et à la diffusion de documentsscientifiques de niveau recherche, publiés ou non,émanant des établissements d’enseignement et derecherche français ou étrangers, des laboratoirespublics ou privés.
Atomization and dispersion of a liquid jet : numericaland experimental approaches
Francisco Felis-Carrasco
To cite this version:Francisco Felis-Carrasco. Atomization and dispersion of a liquid jet : numerical and experimentalapproaches. Fluids mechanics [physics.class-ph]. Ecole Centrale Marseille, 2017. English. NNT :2017ECDM0001. tel-01558141
de vitesse. Une optique de 310mm de distance au plan focal est utilisée comme émetteur,
et de 400mm pour le récepteur. Les deux sont écartées d’un angle de 55°, ce qui permet de
maximiser le taux d’acquisition. L’analyseur de spectre des bouffés Doppler (BSA) est un
modèle P60, également fourni par Dantec-Dynamics.
Les mesures par LDV sont effectuées en deux campagnes différentes : une pour mesurer dans
la phase liquide ; et l’autre pour mesurer dans la phase gazeuse, en utilisant des gouttelettes
d’huile d’olive (d ∼ 1−2µm) comme traceurs (Figure 1b). Dans ce dernier cas, une configura-
tion particulière du BSA permet de différencier la vitesse purement du gaz de celle des gouttes
d’eau qui se trouvent dans le mélange (Mychkovsky et al. [43] [42]).
Les mesures par DTV sont effectuées par une technique d’ombroscopie. Un système Shadow-
Strobe de Dantec-Dynamics est utilisé pour acquérir les images. Le système est composé d’une
caméra CCD PIV/DTV HiSense 4M-C, avec une optique Canon MP-E 65 mm f/2.8, où une
source laser Litron Nd-YAG de 135m J (532nm) est couplée avec un diffuseur/collimateur
qui génère un fond d’image d’intensité lumineuse uniforme et non-cohérente. Des paires
xii
Résumé de la thèse
d’images (frames) sont acquises à une fréquence de fa = 5 H z, où le temps entre chaque frame
est défini entre tbp = 5−20µs , en fonction de la vitesse moyenne des objets dans l’image qui
varie suivant la distance au centre (y/dn = 0−32). La résolution est de 139 pi x/mm, ce qui
correspond à une taille d’image de 14.73×14.73mm2 (2048×2048 pix).
Cette méthode est basée sur les travaux de Yon [65], Fdida et Blaisot [20] et Stevenin et al. [58]
pour l’estimation de tailles de gouttes dans un spray poli-disperse. Les vitesses sont calculées
à partir de l’algorithme SoftAssign proposé par Gold et al. [24]. Cet algorithme a été adapté et
implémenté lors de ces travaux en utilisant le Image Processing Toolbox de MATLAB, à l’aide
d’une carte graphique nVidia CUDA. La Figure 1c montre une photo d’ombroscopie du spray,
où l’algorithme par DTV est appliqué aux gouttes détectées dans l’image. L’estimation des
tailles (contours) et vitesses (vecteurs) y est reportée.
Entrée
Liquide
Sortie
Buse
PMMA
Corps
Verre borosilicate
Capillaire 1.2 mm int.
Acier Inox 8 mm
connecteur rapide
Prise
Capteur de pression
(a) Dessin CAD de l’injecteur. (b) Mesures par LDV. (c) Post-process DTV.
FIGURE 1 – Campagne expérimentale en utilisant une buse de dn = 1.2mm.
Résultats et analyseLes premiers résultats expérimentaux sont analysés pour définir les paramètres de base
caractérisant le comportement des jets. Le premier est l’estimation de la longueur de rupture
du coeur liquide Lc . À partir d’une analyse similaire à celle de Wu et Faeth [63] et Hoyt et Taylor
[29], la Figure 2 met en évidence le régime turbulent de rupture. La valeur expérimentale
calculée à partir de la moyenne des images Lcdn
= 219 est proche de celle estimée à partir de la
relation Lcdn
= 8.51W e0.31L = 203 (Sallam et al. [53]).
FIGURE 2 – Ombroscopie au niveau de l’axe du jet : de x/dn = 0 jusqu’à x/dn = 800.
Dans la zone dispersée du jet (de x/dn = 400 jusqu’à x/dn = 800), des profils radiaux sont
xiii
Résumé de la thèse
acquis par LDV et DTV. L’analyse conjointe des vitesses et tailles de gouttes issue de la
campagne expérimentale est montrée à la Figure 3, où des histogrammes (normalisés en
pdf ) sur un profil radial à x/dn = 800 sont construits. Ces résultats mettent en évidence
l’existence d’une forte vitesse moyenne de glissement entre les deux phases. En plus, en
fonction de leur taille, les gouttes réagissent de façon très différente à la turbulence. À partir
de ces informations, les quantités moyennes décrites dans la Table 1 sont construites pour
comparer aux résultats numériques. Pour la DTV, deux types de moyennes sont calculées :
pondérées ou non par le diamètre des gouttes.
TABLE 1 – Quantités moyennées à partir des données expérimentales.
Méthode Quantité moyennée Formule
LDV-Liq. Vitesse ui ,L = 1n
∑nk=1 ui ,k∈Li q
T. de Reynolds Ri j ,L = 1n
∑nk=1
(ui ,k∈l i q − ui ,L
)(u j ,k∈l i q − u j ,L
)LDV-Gaz Vitesse ui ,G = 1
n
∑nk=1 ui ,k∈Gas
T. de Reynolds Ri j ,G = 1n
∑nk=1
(ui ,k∈g as − ui ,G
)(u j ,k∈g as − u j ,G
)DTV Vitesse ui = 1
n
∑nk=1 ui ,k
T. de Reynolds Ri j = 1n
∑nl=1
(ui ,l − ui
)(u j ,l − u j
)Vitesse pondérée ui ,d =
∑nk=1 d[30],k ui ,k∑n
k=1 d[30],k
T. de Reynolds pondé-
rée
Ri j ,d =∑n
l=1 d[30],l (ui ,l−ui )(u j ,l−u j )∑nl=1 d[30],l
Vitesse par classe ui ,(k) = 1n
∑nl=1 ui ,l∈(k)
T. de Reynolds par
classe
Ri j ,(k) = 1n
∑nl=1
(ui ,l∈(k) − ui ,(k)
)(u j ,l∈(k) − u j ,(k)
)
TABLE 2 – Partition de la population de gouttes par classe de diamètre.
Classe 1 : d[30] ≤ 0.10mm
Classe 2 : 0.10mm < d[30] ≤ 0.25mm
Classe 3 : 0.25mm < d[30] ≤ 0.50mm
Classe 4 : 0.50mm < d[30] ≤ 0.75mm
Classe 5 : 0.75mm < d[30] ≤ 1.00mm
Classe 6 : 1.00mm < d[30]
Les campagnes de mesure par LDV du liquide-gaz sont comparées à celle de la DTV. La façon
de construire les quantités moyennes a une influence sur les résultats dans la représentation
de la phase liquide. Le volume de mesure de la LDV est petit par rapport à l’aire d’intégration
des données de DTV, ce qui la rend plus précise dans l’espace . Pour avoir une précision
supplémentaire dans la DTV, une décomposition en sous-images est effectuée, où les gouttes
détectées sont reparties dans 5 divisions horizontales dans l’image. En plus, pour caractériser
le comportement des gouttes en fonction de leur taille, le classement détaillé dans la Table 2
est utilisé.
xiv
Résumé de la thèse
FIGURE 3 – Histogrammes de vitesses et tailles de gouttes (pdf ) à x/dn = 800. LDV-Gaz, LDV-Liqet DTV.
Les résultats issus, des cas de simulation, des différentes façons de représenter les moyennes
de la DTV et de la LDV, sont présentés à la Figure 4. La portée du jet est caractérisée par
le taux de décroissance de la vitesse sur l’axe. Ruffin et al. [51] ont mis en évidence queux,0
u j= 1
A
(dn
x−x0
)(ρL
ρG
)b, avec b = 0.5 pour un jet gaz-gaz à masse volumique variable, où A ≈ 0.2.
xv
Résumé de la thèse
Ici, on obtient A = 0.021 par LDV et A = 0.019 par DTV, calculés à partir de x/dn > 400.
L’étalement du jet est caractérisé par le paramètre S = ∂y0.5u
∂x , où la demi-largeur de la vitesse
est définie telle que ux,L(x, y = y0.5u) = ux,L,0/2. Ces valeurs (A et S) sont proches de celles
estimées par Stevenin et al. [58] : A = 0.027 et S = 0.024.
0 0.2 0.4 0.6 0.8 1
x (m)
0
10
20
30
40
〈u〉 x
,0(m
/s)
ui: k − ǫ
ui: Rij − ǫ
ui: Rij − ǫij
ui,L: LDV
ui: DTV
dui: DTV
(a) Vitesse axiale (SIM, LDV, DTV).
0 0.2 0.4 0.6 0.8 1
x (m)
0
10
20
30
40
y0.5u(m
m)
S = 0.047
S = 0.030
S = 0.018
S = 0.020
S = 0.021
S = 0.026
(b) Demi-largeur (SIM, LDV, DTV).
0 0.2 0.4 0.6 0.8 1
x (m)
0
10
20
30
40
ux,0(m
/s)
Class 1
Class 2
Class 3
Class 4
Class 5
Class 6
(c) Vitesse axiale par classe de goutte (DTV).
FIGURE 4 – Campagne expérimentale en utilisant une buse de dn = 1.2mm.
Si on considère que la vitesse moyenne calculée à partir de la LDV est la plus précise sur l’axe
du jet, l’écart par classe de goutte observée en DTV, met en évidence que, selon la taille, les
gouttes vont réagir de façon différente à la turbulence de l’écoulement. Le modèle Ri j −ϵi j
semble être le plus proche des résultats expérimentaux. Cette observation est confortée par la
figure 5, où les profils radiaux de vitesse axiale sont comparés. La vitesse axiale de mélange ux
doit être une combinaison de la vitesse de la phase liquide ux,L et du gaz ux,G , en fonction de
la fraction massique Y . Cette dernière quantité est également montrée à la figure 5, mais issue
de la modélisation, comme point référentiel.
xvi
Résumé de la thèse
0 20 40 60 80
y (mm)
0
10
20
30
〈u〉 x
(m/s),Y
(−) k− ǫ : ux
Rij − ǫ : ux
Rij − ǫij : ux
Rij − ǫij : Y (×30)LDV : ux,L
LDV : ux,G
DTV : ux averageDTV : ux d-average
x/dn = 800
FIGURE 5 – Composante axiale de la vitesse moyenne en fonction de la distance radiale.Comparaison des modèles de turbulence vis à vis des résultats de LDV et DTV.
Autant la vitesse moyenne est très bien représentée par la modélisation U-RANS autant les
champs turbulents ne le sont pas. En effet, l’énergie cinétique turbulente est correctement
reproduite, mais sa distribution selon les composantes principales du tenseur de Reynolds
est plus anisotrope que prévue. La Figure 6 montre un comportement très similaire à celui
d’un jet gaz-gaz pour la composante ⟨R⟩11 (voir Hussein et al. [30]). Par contre, la composante
⟨R⟩22 est très faible, avec un facteur d’anisotropie ⟨R⟩22/⟨R⟩11 ≈ 0.05. Ce résultat est similaire à
celui trouvé par Stevenin et al. [58], mais très différent à celui de El-Asrag and Braun [18] dans
un jet d’acétone ou celui de Ferrand et al. [21] dans un jet de gaz avec des particules.
-4 -2 0 2 4
y/y0.5u
0
0.02
0.04
0.06
0.08
0.1
〈R〉 1
1/〈u〉2 x
,0
-4 -2 0 2 4
y/y0.5u
0
0.01
0.02
0.03
0.04
0.05
〈R〉 2
2/〈u〉2 x
,0
-4 -2 0 2 4
y/y0.5u
-0.02
-0.01
0
0.01
0.02
〈R〉 1
2/〈u〉2 x
,0
Rij : k− ǫ
Rij : Rij − ǫ
Rij : Rij − ǫij
Rij,L : LDV
Rij,G : LDV
Rij : DTV
Rij,d : DTV
x/dn = 800
FIGURE 6 – Tenseur de Reynolds en fonction de la distance radiale. Comparaison des modèlesde turbulence vis-à-vis des résultats de LDV et DTV.
En se focalisant sur l’analyse du modèle Ri j −ϵi j , avec une fermeture au second ordre pour
xvii
Résumé de la thèse
u ′′i Y ′′ , une possible source de cette anisotropie peut être la représentation de Σi j (Éq. 9).
En effet, la vitesse de glissement moyenne est directement liée au flux turbulent de masse,
avec ui ,L − ui ,G =u′′
i Y ′′
Y (1−Y ), et pourtant, lié au terme de production Σi j . Par contre, à cause
du gradient de pression, c’est seulement le glissement radial qui intervient de manière
prépondérante. Aussi, Σi j n’est pas en cause. Une possible explication pourrait provenir
du terme de production Pi j . Cependant, ce terme est correctement estimé en fonction de
la composante ⟨R⟩12 et du champ de vitesse ⟨u⟩1. C’est pour cela que nous remettons en
cause le rôle de la redistribution φ(r api d ,Σ)i j qui ne permet pas, dans sa formulation actuelle,
de diminuer la composante ⟨R⟩22 au profit de ⟨R⟩11. Cette dernière hypothèse n’a pas pu être
explorée dans le cadre de ces travaux.
ConclusionsLes points suivants résument les travaux réalisés au cours de cette thèse et ouvrent sur leurs
perspectives :
• Un cas d’étude à échelle réduite est correctement développé pour étudier l’atomisation
d’un jet liquide, dans un régime proche de ceux rencontrés en irrigation et pulvérisation
de pesticides. Les simplifications faites permettent d’assurer une compatibilité entre les
simulations numériques et les mesures expérimentales afin de caractériser finement ce
jet diphasique.
• Un modèle U-RANS de mélange eau/air est implémenté à l’aide des outils CFD Open-
FOAM pour étudier le cas évoqué. La flexibilité du code permet d’explorer correctement
les différents modèles de turbulence et flux turbulent de masse, avec une approche
Eulerienne pour la description de l’interface liquide-gaz. Une stratégie de solution est
proposée dans l’algorithme numérique, ce qui permet d’avoir une solution compatible
avec les équations à masse volumique variable, dans un cas de mélange diphasique
incompressible.
• Pour la campagne expérimentale, des mesures par LDV et DTV sont effectuées. La
mesure de la vitesse par LDV permet d’estimer les champs de vitesse moyenne et
fluctuante dans les deux phases du jet (liquide/gaz). Par ailleurs, la technique de
DTV permet de désagréger l’information du liquide (champs de vitesse moyenne et
fluctuante) par taille de gouttes. L’ensemble de ces données, obtenues par LDV et DTV,
permet de comparer le comportement de ce jet liquide avec les cas de simulation.
• Le comportement dynamique de ce type de jet, décrit par les champs moyens de
vitesse, est très différent des jets monophasiques gaz-gaz. La géométrie et le régime
d’atomisation produisent une faible décroissance de la vitesse sur l’axe et un faible
taux d’étalement du jet. Malgré cela, ces comportements sont bien capturés avec un
modèle de turbulence de type RSM. Sur les champs turbulents, un comportement très
diffèrent selon la taille de gouttes est trouvé pour la contrainte de Reynolds, où le facteur
d’anisotropie peut atteindre ⟨R⟩22/⟨R⟩11 ≈ 0.05. D’un point de vue numérique, cette
anisotropie ne peut pas être bien représentée, ce qui oblige à utiliser un nombre de
xviii
Résumé de la thèse
Schmidt turbulent assez grand pour le flux turbulent de masse σY = 5.5.
• Les perspectives de ces travaux sont évoquées en fonction des améliorations sur la
précision des résultat expérimentaux et l’exploration de nouveaux cas d’étude numé-
riques. Du coté expérimental, des nouveaux systèmes LDV permettraient de faire une
distinction plus précise dans un écoulement diphasique gaz/liquide. L’usage de ce
nouveau système sur une configuration de jet similaire à celle-ci parait pertinente, à la
fois pour valider les résultats obtenus, mais aussi pour tester l’efficacité de cette nouvelle
technique. Les statistiques sur la population de gouttes obtenues par DTV nécessitent
une calibration par rapport à la profondeur de champ. Cette méthode a été mise en
oeuvre dans cette thèse mais les corrélations taille/profondeur de champ et corrections
des contours en fonction des gradients de niveaux de gris n’ont pas été appliquées aux
données DTV. En effet, si la distribution de la population de gouttes est modifiée par la
calibration, les champs de vitesse et de fluctuations doivent l’être également. Aussi, une
telle correction n’est pas triviale et nécessite de plus amples recherches. Finalement,
du coté numérique, une possible source pour augmenter l’anisotropie du tenseur de
Reynolds est proposée : modifier le terme de redistribution dans le modèle turbulence
RSM (φ(r api d ,Σ)i j ) pourrait permettre d’approcher les résultats expérimentaux.
Figure 1.2 – Reynolds stresses anisotropy factor R22/R11 (w ′w ′/u′u′) from the DTV measure-ments performed by Stevenin [57].
This result raises questions about the k − ϵ RANS turbulent model used, and moreover, the
assumptions of a boundary-layer like flow might neglect some key aspects about the source of
this anisotropy.
Indeed, as pointed out in a more recent study by El-Asrag and Braun [18], the use of a RSM
(Reynolds Stress Model) over a k −ϵ model type could improve the prediction of the Reynolds
stresses in zones where the anisotropy is large.
It is then one of the main motivation of this study to find the source of this anisotropy by
investigating why and how it is generated in this type of flow. To achieve this goal, a similar
study case is considered in the present work, where numerical and experimental approaches
are used.
1.1.2 Liquid jet’s fragmentation
The atomization of a liquid jet occurs when a liquid-phase flow is injected into a gas-phase
medium. This two-phase flow is considered non-miscible, meaning that the two phases do
not form a mixture fluid and there are forces that keep a distinguishable interface between
them. By the action of external forces on this interface, the liquid-phase breaks into packets or
droplets, causing the actual atomization into the gas phase.
The forces present in this process of atomization vary depending on the fluid’s properties
and operating conditions. If there is only one liquid phase and one gas phase present, no
phase-change occurs and there are no compressibility effects, the relevant physical properties
are summarised in Table 1.1.
6
1.1. State of the art
Table 1.1 – Physical properties of a phase-incompressible two-phase flow in SI-Units.
ρL Liquid density (kg /m3)
ρG Gas density (kg /m3)
νL Liquid kinematic viscosity (m2/s)
νG Gas kinematic viscosity (m2/s)
σL−G Liquid-Gas surface tension (N /m)
For the operating conditions, in the case of a liquid injected through a nozzle, only the average
bulk velocities of both phases are considered. These are detailed in Table 1.2.
Table 1.2 – Operating conditions of a phase-incompressible two-phase flow.
uL,J Liquid phase average bulk velocity (m/s).
uG ,J Gas phase average bulk velocity (m/s).
Where uG ,J is the injection velocity of a coaxial gas flow. Having these basic physical properties
and operating conditions, three main dimensionless quantities can be constructed as a
function of the forces that intervene in the atomization process:
• Reynolds number: Ratio of inertial forces to viscous forces within a fluid subject move-
ment. Defined at the exit of a nozzle of diameter dn :
Re = (uL − uG )dn
νL. (1.1)
• Weber number: Ratio of inertial forces to surface tension. Can be defined for the liquid:
W eL = ρL(uL − uG )2dn
σL−G, (1.2)
and for the gas:
W eG = ρG (uL − uG )2dn
σL−G. (1.3)
• Ohnesorge number: Relate the viscous forces to inertial and surface tension:
Oh = ρLνL√ρLσL−G dn
. (1.4)
In an extensive review, Dumouchel [16] presents many experimental works on the primary
atomization of liquids. Based on these dimensionless numbers, several classifications can be
made as a function of: fluids properties, geometry, laminar or turbulent regimes, gas assisted
or injected into still gases. In the case of liquid jets for agricultural applications, there is a
7
Chapter 1. General context
high probability to find turbulent liquid round jets. Therefore, the analysis of the atomization
regime is centred on this type of liquid fragmentation.
Having a fixed geometry and working fluid, the average bulk velocity uL is the only parameter
that could set the working regime of a round jet. As detailed by Dumouchel [16], a first
classification can be made based on the observation of the liquid core breakup length Lc as a
function of uL . This is shown in Figure 1.3.
A
B C D E
Figure 1.3 – Round jet behaviour, stability curve of the breakup length Lc as a function ofthe average liquid velocity at the nozzle u J . Region A: Dripping regime. Region B: Rayleighregime. Region C: First wind-induced regime. Region D: Second wind-induced regime. RegionE: Atomization regime. (from Dumouchel [16])
Then, as a function of the Weber and Ohnesorge numbers, several authors described a detailed
separation between the regions as detailed in the Table 1.3.
Table 1.3 – Summary of the criteria for the cylindrical liquid jet fragmentation regimes.
Region A: Dripping regime W eL < 8
Region B: Rayleigh regime W eL > 8
W eG < 0.4 or
W eG < 1.2+3.41Oh0.9
Region C: First wind-induced regime 1.2+3.41Oh0.9 <W eG < 13
Region D: Second wind-induced regime 13 <W eG < 40.3
Region E: Atomization regime 40.3 <W eG
As described by Dumouchel [16], the characteristics of large jets (dn > 1mm) is the presence
of peeling droplets from the nozzle exit, this is called the primary breakup.
Primary breakup is important because it determines the initial properties of the dispersed
8
1.1. State of the art
liquid phase and has an effect on the behaviour of the later secondary breakup mechanism. Wu
et al. [64] showed that spray properties are strongly determined by the turbulence conditions
at the nozzle exit and differ from the results with laminar nozzle conditions. Moreover, the
length of the liquid jet core is also affected by the turbulence inside the injector.
As an example, the main case studied by Stevenin et al. [59] [58] corresponds to a turbulent
high-Weber liquid round jet, whose conditions are summarised on Table 1.4.
Table 1.4 – Experimental conditions used in the study performed by Stevenin [57].
Nozzle diameter dn 4.37 mm
Injection bulk velocity uL 22 m/s
Density ratio ρL/ρG 840
Reynolds number Re 97000
Weber number W eL 29000
Ohnesorge number Oh 0.0018
This would place the case in the Region D of the diagram. Moreover, based on the review by
Sallam et al. [53], the liquid breakup length Lc should follow the following empirical relation:
Lc
dn= 8.51W e0.32
L , (1.5)
corresponding to a turbulent breakup regime, yielding an estimated average breakup length
of Lc /dn = 228. In this regime, breakup is due to the turbulent fluctuations, already present in
the liquid core, leaving the aerodynamic effects to a secondary role.
To study a similar case, whatever the type of round nozzle used, it should operate under the
following considerations:
1. It should be a large circular jet, where dn > 1mm. In a turbulent regime, there should be
a distinguishable boundary layer inside the nozzle, this generates the peeling droplets
at the surface right after the injection.
2. The combination of physical properties of the fluids, along with the geometrical and
operating conditions, should place the atomization regime into the second wind-induced
regime.
This motivates the construction of a specific study case that is carried out throughout this
whole study. Both numerical simulations and experimental techniques are applied to this
study subject, these are detailed later in Chapter 2 and Chapter 3 respectively.
9
Chapter 1. General context
1.2 Study case
As reported by Stevenin [57], the main difficulty for obtaining accurate experimental results
using a DTV and LDV set-up is the spatial precision of the measurement points. When working
with a large liquid jet, the resulting liquid range could be up to several meters, making an
experimental campaign difficult to accomplish.
In an effort to try to reproduce a similar case under a more controlled experimental envi-
ronment, a downsized case is considered. In particular, to investigate the Reynolds stresses
anisotropy as shown before, the downsized study case should be placed in the same atomiza-
tion regime. Considering this, the following parameters for this study case are selected:
1. Injector: A circular nozzle of dn = 1.2mm is used. To avoid any extra difficulty on the
estimation of the boundary layer inside the nozzle, the roughness of the interior walls is
considered negligible. With this, a borosilicate glass is chosen for the material. In the
same way as Wu et al. [64], Sallam et al. [53] and others mentioned in Dumouchel [16]
review on round jets, the nozzle length is chosen in order to obtain a fully developed
turbulent pipe flow, in this case Ln/dn = 50.
2. Fluids properties: A liquid water jet is injected into still air. From this, Table 1.5 shows
the physical properties taken at normal conditions (297 K, 1 atm).
Table 1.5 – Physical properties of the study-case in SI-units at normal conditions.
ρL Water density 998.3 kg /m3
ρG Air density 1.205 kg /m3
νL Water kinematic viscosity 1.004x10−6 m2/s
νG Air kinematic viscosity 15.11x10−6 m2/s
σL−G Water-Air surface tension 0.073 N /m
3. Injection velocity: An injection average bulk velocity of u J = 35m/s is selected. Along
with the physical properties mentioned before, it yields the dimensionless numbers
detailed in Table 1.6.
Table 1.6 – Dimensionless numbers for the study-case conditions.
Reynolds number Re 41833
Weber number W eL 20158
W eG 24.3
Ohnesorge number Oh 0.0034
4. Gravity effects: To avoid any asymmetry, the injection velocity is aligned with gravity,
pointing downwards.
10
1.2. Study case
All these fluids properties and operating conditions ensure that the turbulence inside the
nozzle should be fully developed upon any upstream boundary conditions. Then, the nozzle
diameter is sufficiently large to have a direct influence on the boundary layer thickness inside
the nozzle on the primary atomization. And finally, the experiment should operate inside the
second wind-induced atomization regime.
With the intent to emulate a real case, and although this type of nozzle doest not exist in
any agricultural application, a simplified case like this should provide a more controlled
environment for any experimental and/or numerical simulation.
11
Chapter 1. General context
Summary
The general framework of this study is presented in this chapter. More specifically, the use of
sprinklers in agriculture, like water jets for irrigation or some specific nozzles for pesticides
spraying. From this point-of-view, the following points could summarise this chapter:
• The study of sprinklers in agriculture leads to the study of the atomization process of
liquids. Like in other applications, the flow is almost always turbulent, meaning that the
analysis is centred on the fragmentation of the liquid under turbulent conditions.
• The understanding of this multiphase flow is tackled by experiments and numerical
simulations. The turbulent nature of the flow induces a large spectrum of scales of
motion, making both experimental and numerical studies hard to accomplish.
• A short literature review reveals the advantage of the use of a simplified study case. This
case is finally a water round jet injected into still air. The cylindrical nozzle diameter
is dn = 1.2mm, with a length of Ln/dn = 50. The injection average bulk velocity is
u J = 35m/s, placing the atomization process in a turbulent second-wind induced
regime.
• Similar to previous experimental and numerical studies conducted at IRSTEA Mont-
pellier Centre, LDV and DTV experimental techniques are used to capture the velocity
fields of both liquid and gas phases in the flow. Whereas from the numerical simulation
part, due to the large spatial dimension size of the problem, an Eulerian mixture RANS
turbulence approach is used to simulate the flow.
12
2 Numerical modelling
Introduction
This chapter is dedicated to the numerical modelling of a generic multiphase flow encountered
in a typical liquid jet atomization problem. An Eulerian approach is considered along with a
mixture-fluid variable-density formulation for the liquid/gas mixture. The chapter is divided
in four main sections: Multiphase flow modelling, Turbulence modelling, Numerical solver
and study cases definition.
In Section 2.1 a detailed description of the mixture multiphase formulation is presented. The
transformation from the instantaneous field equations, and their corresponding variables,
to the average mixture problem is achieved using the Favre-average operator. This operator
transforms the set of equations into a variable-density U-RANS (Unsteady Reynolds-averaged
Navier-Stokes) problem.
Based on this formulation, the description of the two main RANS turbulence models used
in this study are presented in Section 2.2: k − ϵ and Ri j − ϵ. Then, several variations for
the turbulent mass flux modelling are presented, along with an Eulerian description of the
interface between the two phases.
The numerical method to solve the U-RANS system of equations is then presented in Section
2.3. Details of the implementation of a custom solver using the OpenFOAM C++ library
are provided. The main focus is on the strong coupling between the turbulent mass flux
and the pressure-based solver in the momentum equation, which differs from a solver for
incompressible constant density fluids.
Finally, several study cases are developed using a combination of the presented models in
this chapter. Starting from the geometric 3D construction of the cases, mesh generation,
convergence analysis and specific definition of every case analysed later in Chapter 4 are
detailed. These cases are constructed based on an incremental analysis on the complexity of
the turbulence and other transport models.
13
Chapter 2. Numerical modelling
2.1 Multiphase flow modelling
2.1.1 Eulerian formulation
For the Eulerian formulation, it is assumed that the fluid is a continuum and the forces applied
to an infinitesimal volume of fluid can be described by field equations. From the starting point
of this case, one main assumption is taken into account: the liquid atomization problem occurs
at high Reynolds and Weber numbers. This means that the forces at the interface between
phases are small compared to inertial forces. This approach yields five main equations for the
instantaneous problem in 3D-Cartesian coordinates (xi with i = 1,2,3):
1. Mass conservation (1 equation):
∂ρ
∂t+ ∂ρui
∂xi= 0; (2.1)
2. Momentum conservation (3 equations, 1 for each component):
∂ρui
∂t+ ∂ρui u j
∂x j=− ∂p
∂xi+ρgi +
∂τi j
∂x j; (2.2)
3. Phase transport (1 equation):
∂ρY
∂t+ ∂ρui Y
∂xi= 0; (2.3)
where every variable is an instantaneous field depending on the absolute position and time
(xi , t ), which in SI units are:
• ui : Velocity field, (m/s).
• p: Pressure field, (Pa).
• gi : Gravity field, (m/s2).
• τi j : Viscous constraint, (kg /m · s).
• Y : Liquid phase indicator, takes the value of 1 when in the liquid and 0 otherwise, (−).
• ρ: Fluid density, takes the value of ρL (liquid density) when in the liquid and ρG (gas
density) otherwise, (kg /m3).
The fluid velocity ui is then composed of discontinuous liquid ui ,L and gas ui ,G velocity fields
at a given position and time. Therefore it is the liquid phase indicator Y which sets the current
state:
ui = Y ui ,L + (1−Y )ui ,G . (2.4)
14
2.1. Multiphase flow modelling
The laminar viscous constraint τi j is modelled using a simple Stokes hypothesis for Newtonian
fluids:
τi j =µ(∂ui
∂x j+ ∂u j
∂xi− 2
3
∂uk
∂xkδi j
), (2.5)
where the dynamic viscosity µ is defined as a discontinuous quantity too, so it takes the value
of µl in the liquid and otherwise µg for the gas as a function of Y :
µ= Y µL + (1−Y )µG . (2.6)
It is important to notice that in this mixture Eulerian formulation for the mixture-fluid there is
no special treatment at the liquid/gas interface, as there are no separate momentum equations
for each phase and the fluid is considered as a miscible binary-mixture. This approach is also
called Quasi-Multiphase Eulerian (QME) in more recent developments [40] [3].
2.1.2 Multiphase average model
Because of the size and the different scales of motion in this liquid atomization problem, a step
further in the modelling involves the averaging of equations (2.1), (2.2) and (2.3); following the
same procedure as in a single-phase variable density fluid [8].
Under this approach a mass-weighted average is used: the Favre Average. It is understood
that the averaging process is an ensemble average over n-identical repetitions, where for any
instantaneous variable h the operator and the subsequent mean h and fluctuating h′′
parts
are:
h = ρh
ρ; h = h +h
′′; (2.7)
where ρ is the mixture density. For a relatively low injection velocity and constant temperature,
ρ is only a function of the mixture of ρL and ρG :
ρ = Y ρL + (1−Y )ρG . (2.8)
The mean volume fraction Y can be expressed also as a function of the mean mass fraction Y ,
making the formulation closer to a variable density scalar mass concentration equivalence:
Y = ρLY
ρ. (2.9)
A graphical representation of this process is shown in Figure 2.1, where the Favre-average
is applied to the liquid phase indicator Y and density ρ, transforming them into Y and ρ
respectively.
15
Chapter 2. Numerical modelling
Figure 2.1 – Favre average operation over the liquid phase indicator Y and the fluid density ρ.
The same procedure can be made to the other variables and the equations (2.1), (2.2) and (2.3),
yielding the desired mixture RANS formulation, which is presented next.
2.1.3 Mixture RANS equations
The mixture RANS model equations are obtained by applying the Favre-average operator to
the previous set of equations and by expressing the variables as a fluctuation centred on the
ensemble average. Using this procedure, the set of equations to solve are very similar to the
previous ones:
1. Mass conservation (1 equation):
∂ρ
∂t+ ∂ρui
∂xi= 0; (2.10)
2. Momentum conservation (3 equations, 1 for each component):
∂ρui
∂t+ ∂ρui u j
∂x j=− ∂p
∂xi+ ρgi +
∂τi j
∂x j−∂ρu ′′
i u′′j
∂x j; (2.11)
3. Turbulent mass transport (1 equation):
∂ρY
∂t+ ∂ρui Y
∂xi=−∂ρ
u ′′i Y ′′
∂xi; (2.12)
where u ′′i u
′′j is the Favre-averaged Reynolds stress tensor and u ′′
i Y ′′ the turbulent mass flux.
Both are new unknowns in the equation and closure models are needed to solve them. The
Favre-averaged laminar viscous constraint τi j is deduced from Eq. (2.5):
τi j = µ(∂ui
∂x j+ ∂u j
∂xi− 2
3
∂uk
∂xkδi j
). (2.13)
16
2.1. Multiphase flow modelling
However, the mixture dynamic viscosity represented by µ is defined using a linear contribution
between the liquid dynamic viscosity µl and gas dynamic viscosity µg :
µ= Y µL + (1−Y )µG . (2.14)
Many forms of this contribution can be found in the literature. For example, Sanjose [54]
used Wilke [62] formulation to describe the mixture viscosity of evaporating fuels, this applies
however only to a mixture of gas species. Despite the inaccuracy of Eq. 2.14, given the
high Reynolds number as a starting hypothesis for this atomization problem, this term is
expected to be orders of magnitude smaller than the Reynolds stresses contribution, making
this possible error negligible.
Two extra expressions arise from this type of averaging. The first one is that the turbulent mass
flux can be expressed from the liquid-gas slip-velocity ui ,S , starting from Eq. (2.4):
ui ,S = ui ,L −ui ,G =u ′′
i Y ′′
Y (1− Y ); (2.15)
where ui ,L and ui ,G are the Reynolds-averaged liquid and gas velocities. The second one is
that the fluctuating part of the Favre-averaged velocity is not centred when a Reynolds average
is applied; u′′i = 0. Indeed, developing this from Eq. (2.7), it can also be expressed in terms ofu ′′
i Y ′′ :
u′′i =−
(1
ρG− 1
ρL
)ρ u ′′
i Y ′′ . (2.16)
Along with momentum and mass conservation equations, and using the same hypothesis for
high Reynolds and Weber numbers flows as Vallet et al. [60], the interface of the liquid/gas
mixture is modelled using a transport equation for the quantity Σ, the mean surface area of
the liquid/gas interface per unit volume.
All variables to solve and quantities to model can be summarised in the following list:
• ρ: Mixture average density (as a function of Y or Y ), to solve.
• ui : Mixture average velocity field, to solve.
• p: Average pressure field, to solve.
• τi j : Mixture average viscous constraint, to be modelled.
• u ′′i u
′′j : Mixture Reynolds stress tensor, to be modelled.
• u ′′i Y ′′ : Turbulent mass flux, to be modelled.
• Σ: Mean surface area of the liquid/gas interface per unit volume, to be modelled. Also
expressed as Σ= ρΩ.
17
Chapter 2. Numerical modelling
The purpose of the next sections of this chapter is to address the solving of this system of
equations.
2.2 Turbulence modelling
The focus of this section is to present all the models implemented and tested into the custom
numerical solver, for both the Reynolds stresses u ′′i u
′′j and turbulent mass fluxes u ′′
i Y ′′ .
2.2.1 Reynolds stresses
Many options exist for the closure of this quantity. However, only two main options are
considered for this study:
1. First order closure: Two-equation variable density k − ϵ model (K-Epsilon).
2. Second order closure: Seven-equation variable density Ri j − ϵ model (RSM, Reynolds
Stress Model).
In the first choice, the Reynolds stresses are coupled with the mean flow using an eddy-viscosity
concept. The form of this eddy-viscosity is then constructed using two transport equations,
both dependent on flow characteristics. The other option is to prescribe transport equations
for each component of the Reynolds stresses and other quantities, also dependent of the flow
characteristics.
First order closure: K-Epsilon
Using an eddy-viscosity model under variable density formulation, a direct transposition from
the Reynolds-averaged Boussinesq hypothesis case is used. Although many variations and
non-linear versions exist for this closure (some can be found fully detailed in Chassaing et al.
[8]), only the simplest linear version is kept.
−ρu ′′i u
′′j +
2
3ρkδi j =µt
(∂ui
∂x j+ ∂u j
∂xi− 2
3
∂uk
∂xkδi j
). (2.17)
Compared to the expression for constant-density incompressible flows, this variable density
version reads that the deviatoric part of u ′′i u
′′j is proportional to the deviatoric part of the
rate-of-strain tensor Si j = 12
(∂ui∂x j
+ ∂u j
∂xi
), via the eddy-viscosity µt , which takes the following
form using a k − ϵ formulation:
µt =Cµρk2
ϵ; (2.18)
18
2.2. Turbulence modelling
where k is the turbulent kinetic energy, ϵ the turbulent kinetic energy dissipation rate and Cµ
a proportional constant. Henceforth, this first order closure centres its efforts into finding
proper transport equations for those quantities.
The exact transport equation for the Favre-averaged turbulent kinetic energy is derived
from the momentum equation Eq. (2.7). The instantaneous values are expressed from the
average and fluctuating parts, then the equation is multiplied by u′′i and averaged, finally the
corresponding summation is applied making k = 12
u ′′i u
′′i . Different versions arises for this
procedure depending on the regrouped parts and their physical explanation [8]. The version
kept is the closest to the later modelled version:
∂ρk
∂t+ ∂ρkui
∂xi (a)
=− ∂
∂x j
[1
2ρãu ′′
i u′′i u
′′j +p ′u
′′j −τi j u
′′i
]
(b)
− ρu ′′i u
′′j
∂ui
∂x j (c)
−τi j∂u
′′i
∂x j (d)
−u′′i
∂p
∂xi (e)
+p ′ ∂u′′i
∂xi ( f )
;
(2.19)
where all the terms in the first row are ones commonly found in constant-density incompress-
ible flows, leaving the second row exclusively to Favre-averaged variable-density flows:
• (a) Material transport in conservative form.
• (b) Diffusion, split in three parts. The first two are the turbulent diffusion, including
pressure effects. The last one correspond to the molecular diffusion. In jet flows, these
two contributions are modelled together using a single gradient diffusion hypothesis:
∂
∂x j
(1
2ρãu ′′
i u′′i u
′′j +p ′u
′′j +τi j u
′′i
)=− ∂
∂x j
[(µ+ µt
σk
)∂k
∂x j
]; with σk = 1.0. (2.20)
• (c) Turbulent kinetic energy production (Pk ) by mean shear, with u ′′i u
′′j from Eq. (2.17).
• (d) Turbulent kinetic energy dissipation rate (Ek ), modelled as ρϵ.
• (e) Energy transfer by coupling the turbulent mass flux with the mean pressure gradient,
also known as the mean pressure work (Σk ).
•(
f)
Pressure-dilatation correlation. It appears when the velocity fluctuation is non-
solenoidal. However, it is not included in the modelled equation.
Consequently, the modelled Favre-averaged variable density k-equation, based on the original
formulation proposed by Jones and Launder [34], is:
∂ρk
∂t+ ∂ρui k
∂xi= ∂
∂x j
[(µ+ µt
σk
)∂k
∂x j
]− ρu ′′
i u′′j
∂ui
∂x j− ρϵ−u
′′i
∂p
∂xi. (2.21)
19
Chapter 2. Numerical modelling
For the turbulent kinetic energy dissipation rate ϵ, a different approach is taken. First, only
the solenoidal part is taken into account, so ϵ∼ ϵ. And second, the modelled equation is not
derived from the exact transport equation for τi j∂u
′′i
∂x j. Instead, an approach is taken in the
same way as Jones and Launder [34] by making the modelled equation homogeneous to the
k-equation counterpart.
Although many options for this modelled equation exist in the literature (an extensive review
can be found in Chassaing et al. [8] and Schiestel [56] for variable density flows), the version
kept is the simplest one and analog to Eq. (2.21):
∂ρϵ
∂t+ ∂ρϵui
∂xi= ∂
∂xi
[(µ+ µt
σϵ
)∂ϵ
∂xi
]−Cϵ1
ϵ
kρu ′′
i u′′j
∂ui
∂x j−Cϵ2ρ
ϵ2
k
+Cϵ3ϵ
kp ′ ∂u
′′k
∂xk−Cϵ4
ϵ
ku
′′i
∂p
∂xi−Cϵ5ρϵ
∂uk
∂xk;
(2.22)
where in the RHS there are in the first row: Diffusion, production, destruction; and in the
second row: the counterparts from Eq. (2.21) of pressure-dilatation and mean pressure work;
being the last one exclusive to compressible flow, related to the turbulence length scale when
passing through a shock-wave.
The standard values for the model constants are Cϵ1 = 1.44 and Cϵ2 = 1.92. The pressure-
dilatation correlation is not modelled, so Cϵ3 = 0. The mean pressure work contribution
counterpart uses Cϵ4 = 1.0. And for the last term, Cϵ5 = 1/3 in isotropic turbulence and
Cϵ5 = 1.0 otherwise (see Chassaing et al. [8, pp. 301-302]). All these parameters are set in
specific study-cases.
Second order closure: RSM
The same strategy as in the previous k − ϵ model is used to define the equations modelled for
the Reynolds stresses. The six equations of the symmetric tensor are extracted from the exact
transport equation for u ′′i u
′′j , whereas the dissipation counterpart is purely modelled.
The base formulation from Launder, Reece, and Rodi [39] is used. As our model aims to
simulate also the flow inside the nozzle, wall-reflexion terms were also included (see Gibson
and Launder [23]). In a similar way to Eq. (2.21), variable density effects were added to the
modelled equation (see Chassaing et al. [8, pp. 312-324]).
The exact transport equation for u ′′i u
′′j , using a specific rearrangement of terms is the following:
∂ρu ′′i u
′′j
∂t+∂ρul
u ′′i u
′′j
∂xl= ρPi j −
∂Tl i j
∂xl+ ρΦi j +Σi j − εi j . (2.23)
In the same way as in the k-equation, some terms need modelled relations to get a complete
20
2.2. Turbulence modelling
closed form equation. A basic linear approach is taken for the construction of these terms,
following the original RSM model from Launder, Reece, and Rodi [39]:
• Pi j , turbulent production. Already in its final form:
Pi j =−( u ′′
i u′′k
∂u j
∂xk+ u ′′
j u′′k
∂ui
∂xk
); (2.24)
• Σi j , Mass flux coupling:
Σi j = u′′i
(∂τ j l
∂xl− ∂p
∂x j
)+u
′′j
(∂τi l
∂xl− ∂p
∂xi
); (2.25)
where only viscous effects are neglected for the modelled part:
Σi j =−u′′i
∂p
∂x j−u
′′j
∂p
∂xi; (2.26)
• Φi j , deviatoric pressure-strain correlation:
Φi j = 1
ρ
⎡⎣p ′(∂u
′′i
∂x j+∂u
′′j
∂xi
)⎤⎦ ; (2.27)
modelled as two contributions, the slow return-to-isotropy Rotta model and the rapid
isotropization of production [47]:
Φi j =φ(sl ow)i j +φ(r api d)
i j ; (2.28)
where:
φ(sl ow)i j =−C1
ϵ
k
(u ′′i u
′′j −
2
3kδi j
); (2.29)
and:
φ(r api d)i j =−C2
(Pi j − 1
3Pl lδi j
)−C3
1
ρ
(Σi j − 1
3Σl lδi j
); (2.30)
where for φ(sl ow)i j , C1 = 1.8, for φ(r api d)
i j , C2 = 0.6 and C3 = 0.75, from Vallet et al. [60].
• Tl i j , transport:
Tl i j = ρãu ′′i u
′′j u
′′l +p ′u
′iδ j l +p ′u
′jδi l −
(τ
′j l u
′′i +τ
′i l u
′′j
); (2.31)
modelled as a whole turbulent diffusion term using the same Reynolds-stress tensor to
21
Chapter 2. Numerical modelling
define an anisotropic diffusion coefficient [47] and the viscous part is neglected:
Tl i j =−Cs ρk
ϵu ′′
l u′′k
∂u ′′i u
′′j
∂xk; (2.32)
where Cs = 0.22.
• εi j , turbulent dissipation rate tensor:
εi j =⎛⎝τ′
j l
∂u′′i
∂xl+τ′
i l
∂u′′j
∂xl
⎞⎠ . (2.33)
The modelled version reads:
εi j = ρϵi j ≡ ρ(
ei j + 2
3ϵδi j
); (2.34)
where ϵ= ϵi i /2 is the turbulent kinetic energy dissipation rate and ei j is the deviatoric
part of ϵi j . Two option are considered. The first one is to neglect the deviatoric part
making ϵi j to act only in the principal components of u ′′i u
′′j :
εi j = 2
3ρϵδi j . (2.35)
The second option is to include some anisotropy as proposed by Rotta [50], but making
the dissipation tensor active in all the components:
εi j = ρu ′′
i u′′j
kϵ (2.36)
This basic model is closer to DNS data in a near-wall boundary layer, but still considered
inaccurate [32]. However, this version is kept and no further analysis is made related to
this type of modelling.
For the kinetic energy dissipation rate ϵ, the same transport equation from the k − ϵ model is
taken. The only main difference is that instead of evaluating the production term using the
Boussinesq relation Eq. (2.17), the explicit Reynolds stresses from Eq. (2.23) are used.
2.2.2 Turbulent mass-flux
Along with the Reynolds stresses, the other main quantity to model is the turbulent mass fluxu ′′i Y ′′ from Eq. (2.12). Given the strong density difference between the liquid/gas, any effect
on the mixture density ρ variation makes the turbulent mass flux strongly coupled with the
whole system of equations in the RANS formulation, and its effect is further transferred into
higher moments via u′′i =−
(1ρG
− 1ρL
)ρ u ′′
i Y ′′ , which is an important source term in the u ′′i u
′′j
22
2.2. Turbulence modelling
transport Eq. (2.23).
To analyse the effect on the behaviour of several case scenarios, three u ′′i Y ′′ closure models are
considered:
• First order model (Mod-0): Basic expression following the gradient of Y and coupled
with turbulence via νt , based on Fick’s law.
• First order model (Mod-1): Basic expression following the gradient of Y but coupled
with the actual Reynolds stresses to include some anisotropy in the behaviour of the
gradient.
• Second order model (Mod-2): A specific transport equation is solved for every compo-
nent of the vector u ′′i Y ′′ , where source terms are coupled with the main flow, turbulence
and an explicit relation to drag forces induced by droplets.
First order model
This approach is similar to a passive scalar transport problem, where if there are no strong
main flow gradients, the concentration of a certain quantity is diffused following a gradient
Fick’s law on itself. As the flow becomes turbulent, the diffusivity coefficient varies, following
the scales of motion in the fluid, but the model is nearly the same.
Based on the original work proposed by Vallet et al. [60], a simplified expression for the
turbulent mass flux was derived by Stevenin et al. [59] by neglecting the pressure gradient
effects and by using a boundary layer approximation on the averaged flow. This approach
ensures that the fluxes are deduced by applying several simplifications on a second-order
model and are not issued as a departure guess:
−ρ u ′′i Y ′′ = µt
σY
∂Y
∂xi; (2.37)
where σY is the turbulent Schmidt number for the diffusivity that takes a value close to 1.0.
However, as experimentally found [58] and assuming a strong anisotropy in a liquid round
jet so that u′′2
2 ≈ 0.082k under the same boundary layer approximation, it yields a value of
σY = 5.5 for the lateral diffusion.
To account dynamically for the possible strong anisotropy in the Reynolds stresses, the
complete approximation of the later expression is the following:
−ρ u ′′i Y ′′ =CY ρ
k
ϵu ′′
i u′′j
∂Y
∂x j. (2.38)
Here, instead of the turbulent viscosity µt , a decomposition using the Reynolds stresses is
used. In the case analysed by Belhadef et al. [5], CY ≈ 0.9, but if the modelled anisotropy is
weak and closer to a mono-phase round-jet, the desired reduction in the lateral diffusivity
23
Chapter 2. Numerical modelling
component (i=2) might not be achieved simply by this type of modelling. Stevenin et al. [59]
proposes a forced way to set this constant in the same way as in Eq. (2.37) model forσY , where
assuming an anisotropy factor u′′2
2/u
′′1
2 = aR such as CY ≈ aRCµ/σY .
Other approaches have been used by other authors to account for this diffusivity variation.
Demoulin et al. [14] tried to makeσY a function of ρ, to account for the large density variations.
Going even further, Desantes et al. [15] proposed a Realizable version of the variable Schmidt
number, by bounding the fluxes with the turbulence fluctuations scales
√u
′′i
2. Nevertheless,
these approaches only change the diffusivity and do not include other effects from the main
flow, which is the purpose of the second order modelling presented next.
Second order model
Although Vallet et al. [60] proposed a second order closure for the turbulent mass flux, this
approach does not provide a direct coupling with the liquid/gas interface surface per unit
volume ρΩ, where the destruction term is only proportional to the turbulence decay rate
τ−1t = ϵ/k.
To tackle this deficiency, a slightly different approach is developed by Beau [4] and later
another similar approach by Andreini et al. [3], who constructed a general framework for the
coupling of ρΩ and Y equations using RANS turbulence models.
The transport equation chosen is the version proposed by Beau [4]. The sink term in this case
is a destruction term by drag forces, induced by the slip velocity between the gas phase and
the droplets:
∂ρ u ′′i Y ′′
∂t+ ∂ρu j
u ′′i Y ′′
∂x j= ∂
∂x j
⎛⎝ µt
σF
∂u ′′i Y ′′
∂x j
⎞⎠−CF 1ρ
u ′′j Y ′′ ∂ui
∂x j−CF 2ρ
u ′′i u
′′j
∂Y
∂x j−CF 3Y ′′ ∂p
∂xi+CF 4F Dr ag ,i ;
(2.39)
where CF 1, CF 2, CF 3 and CF 4 are constants, specified as 1.0, 1.0, 0.0 and 4.0 respectively by
Beau [4], with σF = 0.9 as the turbulent Schmidt number in the diffusion term. The drag force
is calculated using a Schiller-Naumann relation, as a function of the drag coefficient with the
Reynolds number, and the velocity seen by the droplets:
F Dr ag ,i =−18ρGνGY
d 2l
(ui ,L − ui ,G − ui ,D
)(1+0.15Re0.687
d
); (2.40)
where dl is a characteristic length of a droplet population, where for this case the Sauter mean
diameter d32, calculated from the ρΩ solution, is used. The Reynolds number associated with
24
2.2. Turbulence modelling
this diameter is:
Red = ∥ui ,L − ui ,G − ui ,D∥dl
νG; (2.41)
where ui ,D stands for the drift velocity. It is assumed to be the limit at which the velocity
follows a first order model, so using Eq. (2.37) it becomes:
ui ,D = 1
Y (1− Y )
νt
σY
∂Y
∂xi; (2.42)
where σY is specified the same as in the first order model.
With these elements, the droplets relaxation time is defined as:
τR = ρLd 2l
18µG
(1+0.15Re0.687
d
)−1. (2.43)
Since the main objective is to construct a modelled case capable to adapt to a large spectrum
of flow characteristics and geometries, it is expected that this formulation for the turbulent
mass flux u ′′i Y ′′ , coupled with a RSM turbulence model, would give better results than a k − ϵ
using a simple gradient law for the mass fluxes. However, as better discussed in the next
section, the strong coupling of the whole system of equations is particularly challenging to the
numerical solver and not all of the models described here could be used in a straightforward
solution.
2.2.3 Eulerian interface
The last quantity to include in the model is the liquid/gas interface surface per unit volume,
ρΩ or Σ (m2/m3). Ω is constructed so that Ω = Σ/ρ, this is done simply to ensure that the
transport equation can be written in a conservative form.
The model was first proposed by Vallet et al. [60] and it has been subjected to several modifica-
tions in the later years (Beau [4]; Lebas et al. [40]; Duret et al. [17]). It is important to notice that
this type of formulation requires two main assumptions: a high Reynolds number, providing a
strong enough turbulent mixture; and a high Weber number, so the surface tension between
the liquid/gas does not play a significant role at the equilibrium to the described atomization
problem.
Based on the latest advances in this formulation, the latest version proposed by Lebas et al.
[40] is kept, with some considerations taken by Duret et al. [17] based on DNS calculations
used to describe the behaviour of some parameters in the average model. The equation for Ω
25
Chapter 2. Numerical modelling
can be constructed in a conservative form, neglecting the evaporation part:
∂ρΩ
∂t+ ∂ρΩui
∂xi= ∂
∂xi
(µt
σΩ
∂Ω
∂xi
)+Φ (Si ni t +Stur b)+ (1−Φ) (Scol l +S2ndBU ) ; (2.44)
where:
• Φ is a logistic function ([0 1]) that changes the importance of the source terms from the
dense part (Y > 0.5) to the diluted part (Y < 0.1).
• Si ni t is an initialisation term important only in the dense part close to the nozzle.
• Stur b is the production/destruction due to turbulence in the dense part of the spray.
• Scol l is the collision/coalescence source term for the dilute part of the spray.
• S2ndBU is the secondary break-up source term (exclusively) for the dilute part of the
spray.
Because of the lack of information on the construction of such parameters applied to this
study case, only the Stur b term is included inside the model. Then a simplified version of the
equation reads:
∂ρΩ
∂t+ ∂ρΩui
∂xi= ∂
∂xi
(µt
σΩ
∂Ω
∂xi
)+αρΩ
τt
(1− Ω
Ω∗
); (2.45)
where τt = k/ϵ and Ω∗ is the equilibrium value at the smallest scales using an equilibrium
Weber number W e∗ = 1.0:
Ω∗ = 40.5(ρL +ρG )Y (1−Y )k
σL−G ρW e∗. (2.46)
The parameters of the model are set by default, meaning α= 1.0 and σΩ = 1.0.
In the same way that in the original work made by Vallet et al. [60], Ω is linked to the Sauter-
Mean-Diameter d[32] by the following relation:
d[32] = ρLY
ρΩ. (2.47)
2.3 Numerical model
The problem described in the previous section forms a non-linear system of differential
equations. One method to solve them is to cut each equation into small pieces and find
a numerical solution that approaches the real one under some assumptions. Providing a
compatible set of initial and boundary conditions, a finite volume method (FVM) is used to
solve the system obtaining an approximated solution for every variable.
26
2.3. Numerical model
Many available CFD tools offer the capability to find a numerical solution to this type of
problem, ranging from laminar flows to different turbulent and multiphase-turbulent models.
Commonly used commercial solutions offer the possibility to include custom expression to
modelled equations, using user-defined-functions (UDF). However, this approach is always
limited because modifications to the solver core are usually not allowed, making difficult to
solve the actual system of equations and simplifications have to be made to overcome this
situation [5] [38] [37].
In this problem, the system of equations to solve is formed mainly by Eq. (2.10), Eq. (2.11)
and Eq. (2.12). Given the heavy coupling between all the variables, and the intention to solve
it as-is, a custom solver is required to properly model each interaction under a known and
controlled numerical environment.
All the efforts are then redirected to create a custom solver using an open source CFD tool.
For this task, the OpenFOAM® code is chosen. It can handle 3D arbitrary meshes for FVM,
common solvers for the momentum equation are already coded, it includes many dicretization
schemes and mainly because it is supported by a large community working in the same field.
The description of this custom solver is the main subject of this section.
2.3.1 OpenFOAM solver
The OpenFOAM® C++ code was developed by Weller et al. [61] as a free, open-source software
for CFD calculations. Currently it is owned and maintained by the OpenFOAM Foundation1
and distributed exclusively under the General Public Licence (GPL).
Using the equivalent of a module from a commercial CFD software, OpenFOAM® is separated
into specific solvers, each one focused on different physical problems but always sharing a
common library of tools, all following an object-oriented programming in C ++.
Instead of building a study-case using a specific module, the approach here is a little different.
Using a solver from a near-like physics as a baseline, modifications are introduced to it to
meet the specific requirements for the desired physical problem, creating a compiled custom
solver. Applied to this particular problem, one of the main goals of this custom solver is to
find a solution for the coupled system of Eq. (2.10) and Eq. (2.11). To see how this is done in
OpenFOAM, an example on how equations are written and treated in the C ++ code stream is
shown using a laminar case for a single-phase fluid:
∂ρU
∂t+∇·φU −∇·µ∇U =−∇p . (2.48)
This is the momentum conservation equation, where U is the velocity field, φ the mass flux
(simply ρU ), µ the dynamic viscosity and p the pressure field. If this pressure field is known
and some initial and boundary conditions are provided, this differential equation can easily
1The OpenFOAM Foundation: www.openfoam.org
27
Chapter 2. Numerical modelling
be solved using a raw piece of code represented in Figure 2.2:
1 solve(
2 fvm::ddt(rho, U)
3 + fvm::div(phi, U)
4 - fvm::laplacian(mu, U)
5 ==
6 - fvc::grad(p)
7 );
Figure 2.2 – OpenFOAM C++ code to solve the momentum conservation equation.
Related to Eq. (2.48), highlighted in blue are the differential operators, in grey there are two
options, depending if one would like to solve for a variable inside the operator (fvm) or simply
to express the result explicitly (fvc). In this case, the variable to solve is the velocity field U , so
ddt, div and laplacian require implicit discretization schemes for U .
Custom solver strategy
More specific to the multiphase problem treated here, a general strategy to solve the system of
equations could be the following:
1. Solve turbulence mass flux u ′′i Y ′′ and Y .
2. Solve ui and p.
3. Solve turbulence model u ′′i u
′′j .
4. Solve other variables (Σ).
Contrary to the previous single-phase example, the pressure field is generally an unknown,
making the item (2) of the list hard to solve.
Many specific methods exist to solve this system of equations, one of them is the PISO
algorithm (Pressure Implicit with Splitting of Operator) developed by Issa [31], which is
generally well suited for unsteady problems using the smallest amount of computational
resources. However, the convergence of this method under heavy compressible or variable
density flows may not be always assured, for those cases, the SIMPLE (Semi-Implicit Method
for Pressure-Linked Equations) algorithm (Patankar and Spalding [44]) can be used, which
uses under-relaxation factors for both pressure and velocity to stabilise the solution. To
account for this, an hybrid mixing of both algorithms is implemented into OpenFOAM and is
detailed next.
The PIMPLE algorithm implemented in OpenFOAM has been developed by Jasak [33] to solve
the transient momentum equation in conservative form. A brief description on how the
28
2.3. Numerical model
algorithm works is presented here only to describe one of the main modifications to account
for the variable-density mixture multiphase formulation of this problem.
This study-case is considered to be an incompressible problem. A relatively low injection
velocity and pressure-drop inside the injector do not produce compressibility effects, nor
cavitation or phase changes. If both phases stay the same, it is considered to be a phase-
incompressible flow.
A classic PISO solver for solving the transient Navier-Stokes equations for incompressible flows
uses a velocity divergence-free condition to impose mass conservation on each time-step.
However, as the velocity field is actually a mixture velocity u, in variable density this condition
does not meet and ∇· u = 0. Actually, developing Eq. (2.10) gives:
∇· u =− 1
ρ
Dρ
Dt. (2.49)
The PISO algorithm is modified to take into account this effect, where the construction of the
Poisson equation for the pressure solver (detailed next) is derived from the mass conservation
in its complete form, yielding an additional explicit source term in the RHS.
The final correction steps on this modified PISO algorithm work the same as in the original
form, where convergence is checked by mass conservation and pressure solution residuals.
Custom PISO loop
A fully discretized version of momentum equation Eq. (2.11), after all numerical schemes have
been chosen, can be expressed in the following form:
aPU (n+1)P = H
(U (n))−∇p(n+1) (2.50)
where U (n+1)P and p(n+1) are the velocity and pressure fields to solve for, advancing from the
solution in t = t(n) to t = t(n) +∆t = t(n+1). The discretization method yields a matrix-arranged
variables in every cell centre P . The method separates every part of the equation that multiplies
the diagonal elements of the matrix as aP and everything else but the pressure in H(U (n)
).
These operators are both function of the velocity field too, but in the linearization process they
are left behind using the last know solution at t = t(n). For example, if the Reynolds stressesu ′′i u
′′j are included explicitly into the momentum equation, then they are inside the H
(U (n)
)operator, as a function of the previous U (n)
P solution.
Using this decomposition, it is easy to find the solution for the next time-step t = tn+1, simply
dividing by aP :
U (n+1)P = H
(U (n)
)aP
− ∇p(n+1)
aP. (2.51)
29
Chapter 2. Numerical modelling
If the pressure p(n) is used, then the solution is an approximation that would require a
correction, called the momentum predictor. However, an implicit solution for t = t(n+1)
is still preferred, making the pressure p(n+1) still an unknown.
To get both at the same time, the velocity solution U (n+1)P is then injected into the mass
conservation equation to isolate the pressure. To do so, first the cell centred values are
interpolated to cell faces f , creating a flux:
(U (n+1)
P
)f=
(H
(U (n)
)aP
)f
−(∇p(n+1)
aP
)f
. (2.52)
Then, the divergence operator (∇·) is applied to Eq. (2.52), forming the mass conservation Eq.
(2.10). A typical incompressible solver would use ∇·U (n+1)f = 0 as a short form, which is shown
to be not true in variable density flows. Indeed, expanding Eq. (2.10):
∂ui
∂xi=− 1
ρ
Dρ
Dt≈− 1
ρ
∂ρ
∂Y
DY
Dt≡
(1
ρG− 1
ρL
)∂
∂xi
(ρ u ′′
i Y ′′)=−∂u
′′i
∂xi. (2.53)
This imposes the extra constraint to solve the equation, as the divergence of the velocity field
must match the RHS of this expression:
∇·(U (n+1)
P
)f=− 1
ρ
Dρ
Dt; (2.54)
which yields a Poisson equation for the pressure p(n+1):
∇·(
H (U )
aP
)f=∇·
(∇p
ap
)f
+ 1
ρ
Dρ
Dt. (2.55)
The solution of this equation is very time-consuming, taking nearly 80% of the computational
time. Moreover, when using an arbitrary non-orthogonal mesh, several correction steps must
be applied because the pressure gradient is expressed normal to cell faces. This procedure is
done by solving Eq. (2.55) and then re-calculating the corrected gradient each time.
Then, the solution for U (n+1), with p(n+1) known this time, is updated by going back to Eq.
(2.51):
U (n+1)P =U∗
P − ∇p(n+1)
aP; (2.56)
where U∗P is simply
H(U (n))aP
, updated using the last known solution for U (n). Finally, if the
residual of p(n+1) is small enough not to produce further changes to the calculated U (n+1), the
solution has converged.
This original approach should be considered as a simplification of a real compressible solver.
For instance, in the compressible case studied by Payri et al. [45], the pressure equation is
30
2.3. Numerical model
parabolic and the simplification made in Eq. (2.53) does not apply.
Solver global loop
The PISO loop explained before can be repeated several times to achieve convergence inside
the same time-step. However, this provides only a converged solution for ui and p fields,
leaving all other variables behind. To tackle this, the PISO loop is placed inside a global
SIMPLE loop as shown in Figure 2.3, where everything is solved for each time-step ∆t .
Figure 2.3 – Solution control for the customised solver implemented in OpenFOAM for eachtime-step ∆t .
The global loop is solved in the following order for each ∆t :
1. SIMPLE Loop in, solve turbulence mass flux u ′′i Y ′′ and Y : (1). If an extra variable is
needed, the last known converged solution is used (usually from the previous time-step)
as an initial guess.
2. PISO Loop, solve ui and p: (2), (3) and (4). Where ϵoc is the pressure residual for the
orthogonal correction and ϵP for the whole PISO loop. This is important because H and
aP are updated using the new velocity field each time.
3. Solve turbulence model u ′′i u
′′j : (5). The turbulence model is solved using converged
velocity and pressure fields. Turbulence equations include non-negligible explicit source
terms in the RHS, to preserve diagonal-dominance in the iterative solver, turbulence
equations are under-relaxed by a factor of α= 0.5.
4. SIMPLE Loop: Go back to (1) using the calculated solution until the residual ϵS for the
pressure field is small enough. No under-relaxation for the pressure or velocity fields is
needed this time.
5. Solve other variables (Σ): (6). Then go to the next time-step.
As detailed in the analysis on Chapter 4, for a typical study case using 10−8 as a converged
residual, the SIMPLE loop takes 2-3 steps for each time-step, then the PISO loop takes 2 steps
31
Chapter 2. Numerical modelling
for each SIMPLE loop and the OC (Orthogonal Correction) takes 2 for each PISO step. It is
then easy to see why the pressure equation takes most of the time inside the global solver. It is
also this part of the solver which reaches convergence last.
2.3.2 Numerical schemes
Numerical schemes are used to have a linearised and discretized version of every equation
in the system. Momentum conservation Eq. (2.11) is shown to take the form of Eq. (2.50)
assuming that a numerical scheme is used. Then, using simple algebraic matrix operations, a
solution can be found. It is then important to describe how this process is done and why some
selected schemes are chosen to run the case analysis.
The Partial Differential Equation (PDE) system is expressed as derivative operators over
variables both in space and time. In OpenFOAM, every transport equation for a scalar φ
can be expressed as follows:
∂ρφ
∂t (a)
+∇· (ρUφ)
(b)
−∇· (ρΓφ∇φ) (c)
= Sφ(φ
) (d)
; (2.57)
where the terms under brackets are:
• (a): Time derivative.
• (b): Convection.
• (c): Laplacian/Diffusion.
• (d): Linearised source.
The Finite Volume Method (FVM) is based on the integral form of this expression, where Eq.
(2.57) is also satisfied:∫ t+∆t
t
[∂
∂t
∫VP
ρφdV +∫
VP
∇· (ρUφ)
dV −∫
VP
∇· (ρΓφ∇φ)dV
]d t
=∫ t+∆t
t
[∫VP
Sφ(φ
)dV
]d t .
(2.58)
Every term needs a discretization form, first in space and then in time. For this, Figure 2.4
shows the geometric parameters assuming an arbitrary mesh decomposition of a domain
in small volumes, where the interaction of two adjacent volumes of centroids P and N is
represented. VP and VN are the volumes of two adjacent elements, d is the distance between
the centroids, f is the name designation of the face separating the volumes and S f the surface
area vector normal to this face, pointing outwards if the face is considered to be owned by P
as in this case.
32
2.3. Numerical model
Figure 2.4 – Parameters in finite volume discretization (from the OpenFOAM®Programmer’sGuide 2.4.0).
Each term in Eq. (2.58) is then transformed using the interactions from the geometry presented
in Figure 2.4. The details on how this is achieved can be found in any book of numerical
methods for fluid dynamics (eg. [22]) or in the OpenFOAM documentation [26] [25].
As an example, the Laplacian/Diffusion term in Eq. (2.57) is expressed as follows:∫V∇· (Γφ∇φ)
dV =∫
SdS · (Γφ∇φ)=∑
fΓφ f S f ·
(∇φ)f . (2.59)
Then, if the mesh is orthogonal and using the parameters defined in Figure 2.4, an implicit
scheme would read:
S f ·(∇φ)
f = ∥S f ∥φN −φP
∥d∥ , (2.60)
where an algebraic solution for the value of φN can be obtained.
It is important to notice that in this case the diffusivity parameter is linearised (it can also
be a function of φ) and interpolated to cell faces. Then, to have an accurate and robust
discretization scheme, an adequate interpolation method must be used and several passages
to solve the equation might be needed to re-calculate these linearised terms.
First, to get all the expressions in an integral form, in volume and in time, the methods detailed
in Table 2.1 are used.
Table 2.1 – Integration and interpolation methods used in the OpenFOAM solver.
Type Method
Temporal In-
tegration
Euler Implicit/Explicit depending on the discretization
scheme.
Volume Inte-
gration
Gauss Gauss’s theorem of the volume integral for gradi-
ents.
Interpolation Linear Used to pass from cell centres to cell faces.
33
Chapter 2. Numerical modelling
Then, for each type of element in this case study the corresponding spatial discretization
scheme is detailed in Table 2.2.
Table 2.2 – Spatial discretization methods used in the OpenFOAM solver.
Type Method
Temporal
Derivative
Euler Implicit for all temporal derivatives. 1st order
accurate in time.
Convection Upwind Used in every model as a first approximation.
Bounded, 1st order accurate in space.
Limited vanLeer Used in the mass fraction transport. Bounded, 2nd
order accurate in space.
Limited Linear Used for the rest. Bounded/unbounded, 1st/2nd
order accurate in space.
Laplacian Linear Limited Corrected part not greater than 0.5 of the orthogo-
nal part.
Source Linear Implicit When the variable is involved.
Linear Explicit When it is a pure source term.
2.3.3 Mesh and convergence
Mesh construction
The mesh is constructed using the blockMesh utility bundled with OpenFOAM. It creates an
unstructured mesh of hexahedral-type elements transformed from rectangular volumes. A
schematic view of the principal geometric parameters for the mesh construction along with
the boundary conditions is shown in Figure 2.5.
Nozzle wall
Atmospheric Pressure
Inlet:
x
y
z(Wall Thickness)
Figure 2.5 – Schematic representation of the mesh, including boundary conditions.
The Dirichlet-Neumann boundary conditions are specified in the three types of boundary:
34
2.3. Numerical model
Inlet (blue), Nozzle wall (red) and Atmospheric (green). These are detailed for all the variables
using the OpenFOAM representation in Table 2.3.
Table 2.3 – Boundary conditions expressed in OpenFOAM solver.
60X11 transducer separates and conduct the 2C beams. A 310mm focus-length optics is
used for the emitter and 400mm for the receiver, forming a LDV measurement volume of
2.9x0.146x0.146mm3 along the principal directions. A Burst Spectrum Analyser (BSA) P60,
also from Dantec Dynamics, is used to analyse the raw LDV data.
Although the system provided the option for PDA (Phase Doppler Anemometry) measure-
ments, the nature of this liquid round-jet breakup produces highly non-spherical droplets
49
Chapter 3. Experimental campaign
making the PDA data extremely biased. The optical arrangement is configured then to
maximise the LDV data-rate in scattering mode, with a 55° detector angle and no mask in the
receiver optics as shown in Figure 3.3.
LDV-2C
Jet Flow
Detector
y
z
Figure 3.3 – Schematic of the 2-Component LDV set-up for measuring both the liquid and gasphases. The measurement volume size is shown next to the liquid round-jet dimensions forscale.
An extra reduction of the LDV measurement volume is achieved by the 55° detector angle,
where the focusing point cuts the larger dimension as Figure 3.3 shows. This is an important
feature of the set-up, because if the measurement volume is considerably larger than the liquid
jet itself, a lack of detail of the resulting velocity field measurement will be found closer to the
axis. Without this, spatial correlations between two distant internal points, inside the volume,
could be mistaken for timed correlations in the acquired time-series.
Acquisition parameters
The details on the specific configuration for measuring the liquid and gas phases is presented
here. As mentioned before, the goal is to capture the liquid phase velocity field ui ,L and the
gas phase velocity field ui ,G . When measuring only in the liquid phase, the LDV captures the
velocity of the liquid/gas interface of large liquid packets or small droplets. To capture the gas
phase, a second configuration uses olive-oil mist as tracers for the gas around the liquid, where
the processing unit captures the velocity of small droplets of ∼ 1−2µm and liquid droplets.
As described by Mychkovsky et al. [43] [42] in a fluidised bed study, a distinction between the
tracers and the real particles can be made by looking at the Doppler burst signal pedestal. If
one type of particle is considerable bigger than the other, the burst should have a bigger carrier
pedestal too. In a transposition from their case, here the background gas phase is seeded with
very small tracers compared to the poly-dispersed liquid droplets, so the same distinction
should exist.
Other authors have also worked with this technique on bubbly flow, like the ones mentioned
in the review made by Joshi et al. [35], where the main difficulty on this kind of LDV set-up is to
capture a proper Doppler signal from the tracers in the carrier phase, when a heavy dispersed
second phase is present.
50
3.3. Measurement set-up
However, the BSA-P60 from Dantec Dynamics available at IRSTEA Montpellier does not
allow to record the Doppler bust pedestal, as this signal is eliminated from the processor at
the beginning of the burst analysis inside the BSA. With this in mind, a second strategy is
developed by doing two sets of measurements: a first without seeded particles in the gas phase,
and a second with the seeded particles but not in the zones where a large concentration of
water droplets is present.
The two different configurations for the Laser source power (LP), photomultiplier gain (PM),
accepted signal-to-noise ratio (SNR) and band-pass filter (BP) are selected and shown in Table
3.1.
Table 3.1 – LDV BSA set-up for liquid and gas phases analysis.
Configuration Laser Power PM Gain SNR BP-Filter
Water 0.6 W 600 V 4 dB Velocity-span based
Oil 1.1 W 1200 V 8 dB Velocity-span based
Given that the LDV processing module does not allow an actual separation of the signal
acquired in the gas phase configuration (olive-oil particles), some assumptions should be
considered when looking at the gas-phase velocity field.
First, it is noticed that a much greater PM gain is needed to be able to see the olive-oil particles
in the raw Burst-Doppler signal. By increasing this value, along with the desired signal to
noise ratio limit (SNR), yields a big data-rate only for oil droplets, where although large non-
spherical objects have a higher intensity, they are seen much noisier and therefore almost
always rejected by SNR criterion. However, there will always be some droplets that are counted
as part of the gas signal.
Second, because the signal intensity from the water droplets/sacs is higher, the gain in the PM
sensitivity sets a threshold on the positioning of the measurement volume, so no overlapping
between gas and liquid profiles is achieved where a large concentration of water droplets
is present, to avoid damage to the PMs. To better illustrate this, a shadowgraph image is
presented along with the drop-sizer algorithm and DTV post-treatment at x/dn = 400, y/dn = 0
in Figure 3.4. While the Liquid LDV measurements can be made regardless the y axis position,
the LDV on the olive-oil tracers, for the gas phase, can only be made at the left of the orange
line in Figure 3.4.
51
Chapter 3. Experimental campaign
Figure 3.4 – Drop-sizing and DTV post-processing on shadowgraph images at x/dn = 400,y/dn = 0: jet centerline (red), y/dn = 4 mark (orange line), droplets detected (colouredcontours), velocity-vector (blue arrow) and equivalent diameter (d[30] in µm) written next toeach contour detected.
An example of the difference between these two types of measurements is shown in Figure
3.5. Where the bi-variate velocity histograms are shown in a coincident point between the two
profiles, precisely at x/dn = 400, y/dn = 4 (at the orange line in Figure 3.4).
0
10
20
30
40
Ux(m
/s)
-15 -10 -5 0 5 10 15
Uy (m/s)
Water (LIQUID)
0
200
400
600
800
1000
1200
1400
1600
Count
0
10
20
30
40
Ux(m
/s)
-15 -10 -5 0 5 10 15
Uy (m/s)
Olive-Oil (GAS)
0
200
400
600
800
1000
1200
1400
1600
Count
Figure 3.5 – Bi-variate histograms of both velocity components for the liquid phase (blue, left)and the gas phase (green, right). Mean values on red line levels.
52
3.3. Measurement set-up
The left histogram shows the data from the liquid campaign (without tracers). As expected, the
velocity events are concentrated near the jet’s average bulk velocity (uL = 35m/s), pointing
slightly outwards, following the jet’s lateral dispersion.
On the contrary, the right histogram shows the data from the gas campaign. This time, the
velocity distribution has a wider velocity span in both components. The average gas phase
entrainment is seen in the upper part, with a low ux,G , and uy,G < 0, pointing towards the
centre. However, close to the centre line, the added-effect of the large water droplets that
slip into the gas phase analysis can still be seen. By looking at the same region as in the left
histogram, their presence is visible in gas phase measurements.
Nevertheless, given the much larger data-rate of oil droplets, seen by the amount of data
concentrated out of the liquid region in the right histogram, it can be considered that the
whole signal is closer to the gas-phase velocity, and not from a liquid-gas mixture. One aspect
that is not investigated, however, is the correction by resident time of the particles inside the
measurement volume.
As this result shows, a clear distinction between both phases can be constructed in a radial
profile using this LDV method, despite the uncertainty closer to the jet’s axis. It is expected
that the separation between the liquid/gas velocity fields will provide a better insight on the
turbulent quantities, needed for the RANS model described in Chapter 2.
Convergence analysis
Before constructing the profiles, a convergence analysis is made on every averaged quantity,
calculated from the velocity time series on both liquid and gas phases.
The convergence criteria for both cases are set on the calculated Reynolds stresses. Although
the worst case scenario varies from point to point (closer or far from the jet centerline), in
general, the R12 component drives the convergence of every other quantity. Higher order
moments, like Skewness and Flatness could only find a convergence in a reasonable time
closer to the jet’s centre line, where the LDV data rate is high, so they are left out of the global
convergence criteria.
The conditions detailed in Table 3.2 show the minimum requirements to consider a point
converged in the liquid measurements campaign. They are evaluated dynamically at every
acquisition point.
Table 3.2 – Convergence criteria for the LDV liquid points.
Condition Value
Maximum Time 10 min
Maximum number of points 106
Residual on R12 10−4
53
Chapter 3. Experimental campaign
Following these criteria, the convergence on the Reynolds stresses is presented in Figure 3.6
at x/dn = 400, y/dn = 4. It shows the calculated components as a function of the time up
to which the average is calculated. The figure shows that the acquisition stopped at 10 min,
meaning that the residual does not reach 10−4. Despite the lack of convergence, the relative
value is considered enough to represent a point in this case. This becomes more evident when
looking at the real constructed profiles.
0 100 200 300 400 500 600
Time (s)
55
60
65
R11(m
2/s
2)
LDV Liquid: Convergence at x/dn = 400 ; y = 4.8 (mm)
0 100 200 300 400 500 600
Time (s)
3
3.5
R22(m
2/s
2)
0 100 200 300 400 500 600
Time (s)
2.5
3
3.5
R12(m
2/s
2)
Figure 3.6 – Convergence analysis of the Reynolds stresses on the liquid phase at [x/dn = 400,y/dn = 4].
In the same way as the liquid measurement campaign, the gas convergence criteria is sum-
marised in Table 3.3.
Table 3.3 – Convergence criteria for the LDV gas points.
Condition Value
Maximum Time 10 min
Maximum number of points 106
Residual on R12 10−4
As the gas tracer particles give a higher data-rate than liquid droplets alone, the convergence
this time reaches the residual threshold for R12, stopping the acquisition at ∼ 6mi n, as Figure
3.7 shows.
54
3.3. Measurement set-up
0 50 100 150 200 250 300 350 400
Time (s)
60
65
70
R11(m
2/s
2)
LDV Gas: Convergence at x/dn = 400 ; y = 4.8 (mm)
0 50 100 150 200 250 300 350 400
Time (s)
8.5
9
9.5
R22(m
2/s
2)
0 50 100 150 200 250 300 350 400
Time (s)
7
7.5
8
R12(m
2/s
2)
Figure 3.7 – Convergence analysis of the Reynolds stresses on the gas phase at [x/dn = 400,y/dn = 4].
Measurement points
The spatial location of the measured points for the two LDV campaigns is shown in Figure 3.8.
0 200 400 600 800 1000
Spr ading-r te
Re r sent t o
LDV Gas radial pro les
LDV Liquid radial pro les
LDV Liquid axial pro le
Figure 3.8 – Schematic view of the measurement points for the LDV campaign in the studycase.
For the liquid campaign, a first axial profile is acquired continuously from x/dn = 0 to x/dn =800. Then, radial profiles from x/dn = 100 to y/dn = 800 are acquired at a step of ∆x/dn = 100,
following the jet spreading-rate. Because of the strong shading of the liquid jet, only radial
profiles are acquired, from the centerline towards the detector. For a complete lateral profile,
a second detector on the other side would be necessary.
The same procedure for the gas phase campaign is conducted, where radial profiles are
acquired at the same positions. As explained before, the radial gas profiles do no touch the
dense zone of the spray, and therefore they do not touch the centerline. However, they can go
further out into the external zones of the boundary layer around the jet.
55
Chapter 3. Experimental campaign
3.3.2 DTV set-up
Shadow images are used to run a custom DTV algorithm on the dispersed regions of the
spray and to visualise the liquid column primary breakup. These images are generated by the
shadow of the liquid, projected into a high-speed camera, in the presence of a perpendicular
background light. The system is a ShadowStrobe from Dantec Dynamics, mounted as shown
in Figure 3.9.
Laser
Di user
Collimator
Figure 3.9 – Experimental set-up of the DTV system using shadow images.
The system captures two consecutive image frames at high speed, from which the detection
and matching of particles/features can be made. The background lighting is generated by a
double-pulsed laser source, consisting on a Litron Nd-YAG of 135m J (532nm). The light is
then conducted via fibre-optics to a diffuser/collimator, generating a non-coherent uniform
background.
A PIV/DTV HiSense 4M-C CCD camera mounted with a Canon MP-E 65 mm f/2.8 lens is used.
It captures 12-bit depth grey-scale images at 2048x2048 pixels in a double-layer CCD sensor.
With this optical arrangement, the scale resolution is 139 pi x/mm, which transforms in an
image size of 14.73x14.73mm2.
The camera is exposed during the whole duration of a pair of captured frames, leaving to the
laser firing timings control over the exposure times. In between the frames, the first image is
transferred to the sensor’s second layer, leaving the first one ready for the second exposure.
For this reason, images must be taken in a dark room, where the actual exposition time is
∼ 4ns.
The acquisition frequency for a pair of images is set to fa = 5 H z. The time between pulses (tbp)
vary, depending on the average velocity of the objects inside the image. This is an important
parameter to set, because it should be large enough to let the droplets move in-between
frames, but not much so no significant changes to the overall form and/or location pattern of
the objects inside the frame is produced.
As an example, a set of raw shadow images are shown in Figure 3.10, taken at the jet’s centerline,
where tbp = 5µs. Ranging from x/dn = 0 to x/dn = 350, a complete visualisation of the
destabilisation and primary breakup can be seen. Small ligaments can be seen at x/dn = 0
close to the nozzle, as Wu and Faeth [63] explains, they are related to the boundary layer inside
the nozzle, where their sizes are found to be proportional to the turbulent eddies inside the
injector. More downstream, at x/dn = 150 some helical structures can be seen, Hoyt and
56
3.3. Measurement set-up
Taylor [29] explain these structures by the amplification of an helical modal instability, where
aerodynamic effects start to play a more significant role in the turbulent breakup regime.
Figure 3.10 – Shadow images at the jet centerline from x/dn = 0 to x/dn = 800.
Liquid column breakup length
Following the original work by Wu et al. [64] and Sallam et al. [53], a characterisation of the
breakup mechanism can be made, as a function of the breakup length Lc and droplets sizes
d[32]. Moreover, provided that the atomization regime of this study case should be the same as
Stevenin et al. [58] case, the behaviour of such quantities must follow the same relations.
To tackle the first part, and to provide an immediate analysis of the atomization regime, the
mean breakup length Lc is estimated by looking at the breakup events from Figure 3.10. Each
one of these images is taken from a series of 1000 at each point, from where the number of
breakup events are counted at each position.
The reasoning is the following: if at a given point, at a given time, the liquid column presents a
discontinuity, then, there is a probability that the first instantaneous breakup happened at
that point or before, closer to the nozzle. Taking this reasoning to the limit, then the ratio of
the number of events Nb to the total number of images NT , at a given position, should follow
the probability that the average breakup length Lc to be less than or equal to the given position
from the nozzle. The results of this calculation on every set of images from x/dn = 100 to
x/dn = 300 is shown in Figure 3.11.
100 150 200 250 300
x/dn
0
0.2
0.4
0.6
0.8
1
P(L
c≤
x)
Nb/NT
Normal-Distribution Fit
Figure 3.11 – Probability distribution of the liquid column average breakup length Lc as afunction of the normalised distance from the nozzle.
57
Chapter 3. Experimental campaign
As Figure 3.11 shows, the probability that Lc ≤ x fits well a Normal distribution, with a mean
value of Lc /dn = 219. As a reference, the relation given by Sallam et al. [53], based on a best-fit
on several experiments on the same regime gives:
Lc
dn= 8.51W e0.31
L = 203. (3.1)
Shadow image segmentation
Having determined the breakup length, and the distance from the nozzle at which liquid
column does not exist anymore, the secondary breakup of liquid packets/droplets begins.
From x/dn = 400 to x/dn = 800, a DTV (Droplet Tracking Velocimetry) algorithm is used to
characterise the dispersed part of the flow.
The main objective is to accurately detect the droplets in a shadow image and calculate their
velocity using the double-frame acquisition. To do so, several image treatment techniques
are used to filter and segment the shadow images. To achieve this, a custom shadow-sizer
algorithm is developed and implemented using the Image Processing Toolbox in MATLAB®.
This shadow-sizer software is an extension to the one developed by Stevenin [57].
To take advantage of the parallel computation capabilities, the new version of the code runs
in parallel, on every computational core available on a x86-64 PC. In addition to that, an
nVidia CUDA enabled graphics-card (Maxwell architecture) is used to perform heavy matrix-
operations, like filtering, bi-linear interpolations and binary operations.
A step-by-step procedure on the general aspects of the code is detailed next:
• Wavelet transform: Based on the procedure presented by Yon [65], a Mexican-hat kernel
function is used and applied as a filter to the original image (Imor g ), the goal is to detect
changes of the image gradient, therefore, amplifying the boundaries of the droplets
no-matter the defocusing (Imw t ). Then, a dynamic threshold generates a binary image
(ImBW ). Every object detected is then a candidate to be a droplet, as Figure 3.12 shows,
where the coloured BW objects are shown to clearly identify the segmentation result.
(a) Original Image (b) Wavelet transform (c) Segmented binary
• Local analysis: Following the local analysis detailed by Fdida and Blaisot [20], every
object is isolated and analysed locally. A local image is created for every object, by
applying the binary mask of Figure 3.12-(c) to Imor g , resulting in a subset of smaller
images Iml oc (local image).
The grey-level intensity of the local images are defined as i , where imi n and imax are
the minimum (dark) and maximum (bright) values. Then, the contrast ratio defined as
C = (imax − imi n)/(imax + imi n) is calculated and the object is rejected from the analysis
if C < 0.1.
• Contours extraction: The objects that pass the contrast filter are finally analysed. The
Iml oc is normalised, meaning that the global grey levels from before now are 0 < i < 1.
From these normalised local images, the contours (w) at the following grey-levels (l ) are
extracted: wl=0.25, wl=0.50, wl=0.61 and wl=0.77. Using a bi-linear interpolation, those
contours are represented in a sub-pixel domain at 10x the original size. Finally, using
the 3D representation described by Daves et al. [10], the equivalent diameter d[30] is
calculated. Other quantities are also kept for further analysis, like the principal axis,
orientation and eccentricity. These are shown in Figure 3.13.
Figure 3.13 – Sub-image analysis on a detected droplet. Local contours and principal axis.
• Velocity estimation: Using the centroids from every droplet detected, on (x, y) coordi-
nates in the pair of images, the SoftAsign matching algorithm form Gold et al. [24] is used.
This creates an output matrix with the matched objects from both consecutive frames.
Finally, knowing the scale resolution (139 pi x/mm) and the time between images (tbp),
the velocity vector of every droplet can be estimated.
The final result of the image segmentation procedure (wavelet transform and filtering, local
analysis, contours extraction and velocity estimation) is shown in Figure 3.14. The information
for every frame is kept for further granulometry and velocimetry of the jet.
59
Chapter 3. Experimental campaign
(a) Superposition of two consecutive shadow 12-bit images using a tbp = 5µs.
0.88 1 mmU = 40 m/s
V = 40 m/s
(b) Shadow-sizer and DTV post-processing. Con-tours at l = 61% and velocity vectors.
Figure 3.14 – Custom DTV post-processing algorithm. Image centre at x/dn = 600, y/dn = 0.
Depth-of-field calibration
Although the Shadow-sizer algorithm can well detect out-of-focus droplets, as a function of
their characteristic sizes (d[30]), these are not always in the same plane of measurement. In
a jet with cylindrical geometry, aggregating information that does not exist within the same
physical space could lead to several biases in the granulometry and velocimetry.
A calibration procedure on the size of the detected objects is then conducted, following
the original work by Fdida and Blaisot [20]. The response of this optical system is studied
by looking at the in-focus and out-of-focus images on calibrated opaque disks of known
size. Despite that droplets are transparent, the refractive index change at the surface does
not influence the grey-level gradient detected at the edges of an opaque object (Fdida [19]),
making this type of calibration on completely opaque objects well suited for droplets.
60
3.3. Measurement set-up
Figure 3.15 – ThorLabs grid distortion target. 3in x 3in, 125 to 2000 µm grid spacings, sodalime glass.
The procedure is then to use the same shadow-sizer algorithm to detect the objects contours,
as a function of a known distance from the focus plane z. The calibrated objects are from
a Thor-Labs grid distortion target (Figure 3.15), where 62.5, 125, 250, 500 and 1000 µm low-
reflexion discs are painted in a soda lime glass support. The results are shown in Figure 3.16,
as a function of the normalised contrast ratio C0.
-20 -10 0 10 20
z (mm)
0
0.2
0.4
0.6
0.8
1
C0
(a) Normalized contrast ratio C0 as a function ofthe distance from the focal plane z for all discdiameters in the target.
0 0.2 0.4 0.6 0.8 1
C0
0
0.2
0.4
0.6
0.8
1
d0/d
m
1000µm500µm250µm125µm62.5µmFit
(b) Ratio of the measured diameter dm at l = 61%to the actual diameter d0 as a function of thenormalized contrast ratio C0.
Figure 3.16 – Calibration using a commercial optical target.
Figure 3.16a shows the normalised contrast ratio C0 (C , where Cmax = 1), for every disc real
size, against the distance from the focal plane z. A difference on the response to the focal
plane distance can be seen for every type of object, this generates a Depth-of-Field (DOF) as a
function of the size of the object.
As also shown by Fdida and Blaisot [20], Figure 3.16b shows the ratio of the actual size of the
objects d0 to the measured dm at l = 61% ((d[30],l=61%) against the normalised contrast ratio C0.
61
Chapter 3. Experimental campaign
From this, it can be seen that the overestimation of the real size follows the same relation no
matter the original size nor the out-of-focus distance and it is only a function of the contrast
ratio.
Following this analysis, every equivalent diameter d[30], extracted from the contours at l = 61%,
for every droplet detected in this study case, can be corrected only by looking at the calculated
contrast ratio C .
This calibration methodology has been developed and tested on cases with nearly round
objects. Here, however, its use is questionable, where heavy deformed large packets of liquid
can be found and the definition of an equivalent diameter is only referential. These corrections
are finally not applied to the results presented in Chapter 4, but they are kept as a reference
about the response of the optical system. This notion of DOF is useful to give a proper
interpretation for the velocity and fluctuations fields presented in Chapter 4.
Convergence analysis
With the analysis on the images in mind, where the droplets sizes and velocities are extracted,
the resulting long series of data are used to construct averaged quantities. In a similar way as
in the LDV case, the average velocity field of these droplets are represented in a spatial grid.
An analysis on how this average representation behaves is presented next.
To construct the average fields, a convergence analysis is first performed as a function of
the number of images (N ) needed to have representative average fields. A previous work
performed Stevenin [57] shows that the joint distribution of droplets sizes and velocities
is very sparse for a similar liquid jet, meaning that the average velocity field has a strong
dependence on the droplets distribution. It is therefore important to study how many objects
are detected and validated in each pair of frames, for when the averaged velocity and droplet’s
size fields are constructed, those quantities seem converged on a N number of total images.
Subsequently, an analysis based on a specific distribution of droplets by class of diameter
is proposed. The aim is to specify a decomposition by sizes where the average fields are
calculated. To do so, the following parameters are set:
1. The partition should be the same for the whole analysis. It is known that large droplets
will exist only close to the centerline and will tend to disappear in the outside regions.
The partition proposed must not change according to this, and if large droplets do not
exist at one point, the class is considered non-existent.
2. The distribution should be minimal. Meaning that a partition of many classes that has
the same behaviour of a smaller one is discarded.
3. The average quantity must be independent if weighted by the diameter inside the class
62
3.3. Measurement set-up
(i ). Meaning that, if h is the quantity to average, then:
h = dh
d. (3.2)
4. The number of elements should allow a convergence of the average inside the class.
5. To avoid loners, a minimum of 100 elements is allowed inside a class. If not, the class is
considered non existent on that point. This is imposed principally because the average
on large droplets will never converge outside the jet centre line, despite that some rare
events occur.
Following these directives, the following partition by droplet equivalent diameter class is
proposed in Table 3.4.
Table 3.4 – Partition of droplets population by class of diameter.
Class 1: d[30] ≤ 0.10mm
Class 2: 0.10mm < d[30] ≤ 0.25mm
Class 3: 0.25mm < d[30] ≤ 0.50mm
Class 4: 0.50mm < d[30] ≤ 0.75mm
Class 5: 0.75mm < d[30] ≤ 1.00mm
Class 6: 1.00mm < d[30]
To justify the use of this partition, the analysis is presented at x/dn = 600, y/dn = 0 (see Figure
3.14), on a set of 1000 images. The data collected corresponds to a 1/5th of the image in the
central point, using the sub-image partition.
The averaging procedure is the following. Average is the mean velocity component ui , calcu-
lated from the arithmetic average, over n objects (droplets) identified as j inside the class (k),
ui ,(k) =1
n
n∑j=1
ui , j∈(k); (3.3)
and d-Average is the same mean velocity component ui , but calculated from the weighted
average, over the same objects, by the droplet d[30] diameter, meaning:
ui ,(k) =∑n
j=1 d[30], j∈(k)ui , j∈(k)∑nj=1 d[30], j∈(k)
. (3.4)
Using this, the influence of the droplet sizes, inside a class, is weighted in the mean velocity
estimation. The convergence is shown in Figure 3.17 for the velocity field.
63
Chapter 3. Experimental campaign
0 2000 4000
Number of droplets
0
10
20
30
40
uxm/s
-1
-0.5
0
0.5
1
uym/s
Class: d[30] < 0.100 mm
0 2000 4000 6000
Number of droplets
0
10
20
30
40
uxm/s
-1
-0.5
0
0.5
1
uym/s
Class: 0.100 mm < d[30] < 0.250 mm
0 1000 2000 3000
Number of droplets
0
10
20
30
40
uxm/s
-1
-0.5
0
0.5
1
uym/s
Class: 0.250 mm < d[30] < 0.500 mm
Average
d-Average
0 500 1000
Number of droplets
0
10
20
30
40
uxm/s
-1
-0.5
0
0.5
1
uym/s
Class: 0.500 mm < d[30] < 0.750 mm
0 200 400 600
Number of droplets
0
10
20
30
40uxm/s
-1
-0.5
0
0.5
1
uym/s
Class: 0.750 mm < d[30] < 1.000 mm
0 500 1000
Number of droplets
0
10
20
30
40
uxm/s
-1
-0.5
0
0.5
1
uym/s
Class: 1.000 mm < d[30]
x/dn = 600 — y = 0.0 mm
Figure 3.17 – Convergence analysis on the mean velocity by droplet’s class diameters. Sub-image count at [x/dn = 600, y = 0mm].
To summarise, this analysis shows that 1000 images are enough to represent the average
velocity field, under this partition by class of droplet diameter, on a sub-image of a 1/5th of
the original lateral size.
However, this is not always true for the calculated Reynolds stresses. These are calculated
relative to the average by class shown before. Figure 3.18 shows the same analysis on the
principal Reynolds stresses by class, R(i )11 and R(i )
22 .
The same principle apply for the Reynolds stresses calculation. Average uses the arithmetic
average:
Ri j ,(k) =1
n
n∑l=1
(ui ,l∈(k) − ui ,(k)
)(u j ,l∈(k) − u j ,(k)
); (3.5)
and d-Average is calculated as a weighted average by the droplet d[30] diameter, meaning:
Ri j ,(k) =∑n
l=1 d[30], j∈(k)(ui ,l∈(k) − ui ,(k)
)(u j ,l∈(k) − u j ,(k)
)∑nl=1 d[30],l∈(k)
. (3.6)
Since the population of large droplets is considerably lower in the bigger class, the convergence
on the number of droplets needed does not always meet. Moreover, there is a difference this
time as a function of the weighted average inside the class, as shown by the doted lines against
the continuous one.
Despite all these difficulties the partition by class is kept and the number of images is not
64
3.3. Measurement set-up
modified either. As the later results show, the extra precision that can be gained by re-setting
those parameters would not change the analysis.
0 2000 4000
Number of droplets
0
10
20
30
40
50
R11(m
2/s
2)
0
1
2
3
4
5
R22(m
2/s
2)
Class: d[30] < 0.100 mm
0 2000 4000 6000
Number of droplets
0
10
20
30
40
50
R11(m
2/s
2)
0
1
2
3
4
5
R22(m
2/s
2)
Class: 0.100 mm < d[30] < 0.250 mm
Average
d-Average
0 1000 2000 3000
Number of droplets
0
5
10
R11(m
2/s
2)
0
0.5
1
R22(m
2/s
2)
Class: 0.250 mm < d[30] < 0.500 mm
0 500 1000
Number of droplets
0
5
10
R11(m
2/s
2)
0
0.5
1
R22(m
2/s
2)
Class: 0.500 mm < d[30] < 0.750 mm
0 200 400 600
Number of droplets
0
1
2
3
4
5
R11(m
2/s
2)
0
0.1
0.2
0.3
0.4
R22(m
2/s
2)
Class: 0.750 mm < d[30] < 1.000 mm
0 500 1000
Number of droplets
0
1
2
3
4
5
R11(m
2/s
2)
0
0.1
0.2
0.3
0.4
R22(m
2/s
2)
Class: 1.000 mm < d[30]
x/dn = 600 — y = 0.0 mm
Figure 3.18 – Convergence analysis on the Reynolds stresses by droplet’s class diameters.Sub-image count at [x/dn = 600, y = 0mm].
Measurement points
As detailed before, all statistics are in a spatial partition of the original image. Since the spatial
resolution is rather big (139 pix/mm; 14.73 mm), there are strong gradients of any quantity if
represented in a lateral profile inside a 2048x2048 pixels frame. To show this effect, the joint
probability density pd f of all events ux -uy , ux -d[30] and uy -d[30] in a sub-image analysis is
presented in Figure 3.19.
This shows that the distributions of velocities and sizes vary a lot inside the image itself. These
points are represented centred on the slices (see Figure 3.20). The probability density functions
(pd f ) are constructed from the histograms. A partition of 50 bins is used to do the count on
every axis, where the white bins are the zones with zero counts. The solid lines represent the
average values and the dashed ones the standard deviation.
Then, the reconstruction of radial profiles is performed by considering only the information
extracted at the central slices, like Figure 3.20 shows. This is called a super-resolution profile,
giving an extra spatial precision on the averaged quantities.
The images are acquired laterally with a step of 4.8 mm, meaning that there is a large overlap-
ping of information. This produces a good quality profile, with enough resolution to perform
further operations, like spatial derivatives.
65
Chapter 3. Experimental campaign
0
10
20
30
40
ux(m
/s)
-5 0 5
y = −5.94 mm
0
10
20
30
40
-5 0 5
y = −2.97 mm
0
10
20
30
40
-5 0 5
uy (m/s)
y = 0.00 mm
0
10
20
30
40
-5 0 5
y = 2.97 mm
0
10
20
30
40
-5 0 5
y = 5.94 mm
0
10
20
30
40
ux(m
/s)
0.5 1 1.5 2
0
10
20
30
40
0.5 1 1.5 2
0
10
20
30
40
0.5 1 1.5 2
d30 (mm)
0
10
20
30
40
0.5 1 1.5 2
0
10
20
30
40
0.5 1 1.5 2
-5
0
5
uy(m
/s)
0.5 1 1.5 2
-5
0
5
0.5 1 1.5 2
-5
0
5
0.5 1 1.5 2
d30 (mm)
-5
0
5
0.5 1 1.5 2
-5
0
5
0.5 1 1.5 2
Figure 3.19 – Bi-variate histograms normalised as a pdf for: ux -uy , ux -d[30] and uy -d[30].Sub-image count at [x/dn = 600, y = 0mm].
66
3.3. Measurement set-up
Figure 3.20 – Super-resolution profile reconstruction by overlapping of sub-image data. Bluezones are kept, red are discarded. Example at x/dn = 600.
Finally, similar to the LDV case, this process is repeated on the locations detailed in Figure
3.21, completing the DTV measurement campaign.
0 200 400 600 800 1000
Spr ading-r te
Re r sent t o
DTV radial profiles
Figure 3.21 – Schematic view of the measurement points for the DTV campaign.
67
Chapter 3. Experimental campaign
Summary
A detailed description of the experimental techniques applied to a study case is given in this
chapter. From this development, the following points could summarise the topics treated:
• LDV and DTV measurements systems are chosen to carry out the experimental cam-
paign. Both systems have been tested before in similar cases by other authors. The
objective is to capture the velocity fields of both gas and liquid phases, along with the
distribution of droplets sizes in the dispersed region of the flow. Additionally, shadow
images taken close to the nozzle allow a visualisation of the liquid column breakup
behaviour.
• The study case consists in a circular glass tube nozzle of dn = 1.2mm. This geometry
allows a direct equivalence with the simulation cases in Chapter 2. Although this nozzle
does not exist in any real application, this simplified case provides a better controlled
environment to perform the experiments, with less incertitude on the operating condi-
tions.
• The LDV technique is applied to measure both liquid and gas phases. These measure-
ments are carried out in separate experimental campaigns. The liquid campaign is
performed in the dense and dispersed part of the jet. For the gas campaign, olive-oil
mist is used as passive tracers in the surrounding air. A special set-up in the acquisition
parameters of the LDV (BP filter, SNR, PM gain) is used to discriminate the average
signal from the tracers and the residual part from the liquid droplets. This configuration
allows to capture the average and fluctuating components of the liquid/gas slip-velocity,
an important quantity to compare with the U-RANS model.
• Shadow images are acquired to run a custom DTV algorithm. From this technique,
the droplets sizes and velocity distribution are obtained, a more detailed piece of
information than the one inferred from the average liquid LDV. A strong relation of
the droplets distribution with both average and turbulent velocity fields is found. This
means that a correct estimation of the DOF is crucial to get an accurate velocity field. To
tackle this incertitude, a calibration procedure is carried out using a calibrated target.
Some of the experimental results are shown in this chapter as an example. This is done to
show how the set-up is done and how the raw data from the LDV and DTV are integrated into
the construction of the averaged fields.
68
4 Results
Introduction
On this chapter, the results from the simulation cases are presented along with the experimen-
tal observations all-together. The development of this work does not follow the same logic
separation of the numerical and experimental chapters. Indeed, the results are presented in a
way so that the mixture multiphase flow model used here could be compared and explained
with and by the experimental observations.
The experimental results are presented first. Their analysis allows to characterise and to set
some overall parameters on the dynamics of the studied liquid jet. The decay of the centerline
velocity or the spreading-rate on a round-jet are some immediate useful parameters to look-up
to, these set the first baseline to quickly compare against numerical simulations. Later, the
mean and fluctuating velocity fields are obtained from both LDV and DTV measurement
campaigns, these quantities are useful to analyse the behaviour of the turbulence RANS model
used [47][8][56].
Great effort also is put into the numerical simulations, where the custom solver is constructed
and implemented using the OpenFOAM CFD code. Although the construction and test of this
solver could be a subject on its own, based on the experimental observations, a series of study
cases are created to test the behaviour of such formulations, which are only applied to this
study case. Always centred on the same Favre-averaged mixture multiphase modelling, some
variations of k−ϵ and RSM turbulence models are compared, along with first and second-order
closures for the modelled turbulent mass transport fluxes [60].
As previously mentioned, one of the main challenges of this work is to find an explanation to
the strong anisotropy found on the Reynolds stresses principal components [58]. A combined
approach, from the experimental observations, seen by the RSM turbulence model, could
provide some clues on this behaviour.
69
Chapter 4. Results
4.1 Experimental observations
The experimental observations are based on the data provided by the LDV and DTV cam-
paigns. From these data, two main quantities are calculated: the average velocity field and
the Reynolds stresses. Some special particularities are involved when performing these
calculations, like the ones detailed in Chapter 3. The objective is to extract the averaged
fields from the liquid and gas phases separately.
In this section, the detailed results on the averaged fields are presented. A comparison between
the two measurement techniques in terms of the average velocity fields and the Reynolds
stresses is attempted first. The details on how these quantities evolve in comparison with the
RANS model are described later on a series of simulation cases.
4.1.1 Mean velocity field
Given the cylindrical symmetry of the flow, there are only two main components involved in
the velocity field: ux and uy , this last one similar to the radial component in a symmetry plane.
No matter what type of average operator is used, the flow is always considered statistically
axisymmetric.
As previously defined in Chapter 3, the averaging procedure differs slightly from the two
analysis. The LDV data is always treated using a simple arithmetic average, over n events,
separated by phase, meaning:
ui ,L = 1
n
n∑k=1
ui ,k∈Li q; (4.1)
ui ,G = 1
n
n∑k=1
ui ,k∈Gas. (4.2)
For the DTV, the procedure is slightly different. The first approach is to calculate the mean
velocities using a simple arithmetic average, like in the LDV case, meaning that the mean
velocity flagged as Average is:
ui = 1
n
n∑k=1
ui ,k ; (4.3)
whereas d-Average is the same mean velocity component ui , but calculated from the weighted
average, over the same objects, by the droplet d[30] diameter, meaning:
ui =∑n
k=1 d[30],k ui ,k∑nk=1 d[30],k
. (4.4)
If the DTV data is separated by class, then a simple arithmetic average is kept, meaning that
70
4.1. Experimental observations
the average velocity for the class (k) is:
ui ,(k) =1
n
n∑l=1
ui ,l∈(k); (4.5)
The centerline axial velocity is then defined as ux,L,0 = ux,L(x, y = 0), where the underscore
indicates the axis component and phase, and in parenthesis the position. From this, the axial
velocity half-width y0.5u is the distance from the jet centerline at which ux,L(x, y = y0.5u) =ux,L,0/2. Using this, the spreading rate S is defined as:
S = ∂y0.5u
∂x; (4.6)
Despite the formal definition of S, it is calculated and considered linear for x/dn > 400.
Using these definitions, Figure 4.1 shows the decay of ux,L,0 against the axial distance from the
nozzle, along with the axial velocity half-width y0.5u for the two measurements techniques.
0 0.2 0.4 0.6 0.8 1 1.2
x (m)
0
10
20
30
40
ux,L,0(m
/s)
0
0.005
0.01
0.015
0.02
0.025
y 0.5u(m
)
Velocity
Half-width (S = 0.020)
(a) LDV. Mean liquid axial velocity along thecenterline uL,x,0 and half-width y0.5u from radialprofiles.
0 0.2 0.4 0.6 0.8 1 1.2
x (m)
0
10
20
30
40
ux,0(m
/s)
0
0.005
0.01
0.015
0.02
0.025
y 0.5u(m
)
Averaged-AverageAverage (S = 0.021)d-Average (S = 0.026)
(b) DTV. Mean liquid axial velocity along thecenterline, ux,0 and diameter-weighted dux,0/d0;and half-width y0.5u from radial profiles.
Figure 4.1 – Mean velocity axial profiles from experimental observations.
In the same way as in constant density round-jets (air-air of Hussein et al. [30]), the jet
looses velocity and spreads. Figure 4.1a shows the decay of uL,x,0 and the spreading of the
jet, characterised by the y0.5u , using the LDV data on the positions defined in Figure 3.8. To
compare, Figure 4.1b shows the same, but using the two averaging procedures proposed from
the DTV profiles.
The difference between the two sets of measurements is explained by the integration volume
on which the data is acquired. In one hand, the dimensions of the LDV measurement volume
are 0.146x0.146x2.9mm3 (see Figure 3.3), making the spatial velocities integration to around
the size of the jet original diameter. On the other hand, the DTV central slice (see Figure 3.20)
is wide 2.9mm too, making both comparable. However, the depth-of-field (DOF) of the DTV
data has a much larger span, even for small droplets (see Figure 3.16a), meaning that the
calculated DTV average of droplets velocities is integrated into a much larger volume than in
71
Chapter 4. Results
the LDV case. Since the axial velocity decays against the radial distance, this larger integration
volume makes the DTV centerline velocity lower than in the liquid LDV case.
Although the larger droplets are seen in a larger DOF, they only exist in the central portion of the
jet, and as previously seen in Figure 3.19, they tend to keep the jet bulk velocity (u J = 35m/s).
This effect is shown by calculating the centerline velocity by class of droplet in Figure 4.2.
0 0.2 0.4 0.6 0.8 1
x (m)
0
10
20
30
40
ux,0(m
/s)
Class: d[30] < 0.100 mm
Class: 0.100 mm < d[30] < 0.250 mm
Class: 0.250 mm < d[30] < 0.500 mm
Class: 0.500 mm < d[30] < 0.750 mm
Class: 0.750 mm < d[30] < 1.000 mm
Class: 1.000 mm < d[30]
Figure 4.2 – DTV. Mean liquid axial velocity along the centerline ux,0 by droplets classdiameters.
From the difference on these results, it is not straightforward to assimilate the LDV or DTV
data to the Liquid phase velocity field. Nevertheless, this is carried forward to complete the
analysis.
Contrary to a constant density case, because of the extra inertial, viscous and gravity effects on
the mixture, in variable density flows there is no straightforward similarity. However, similar
relations can be found in the literature. Ruffin et al. [51] studied the decay rate of several
variable density flows, where the following relation can be applied:
ux,0
u J= 1
A
(dn
x −x0
)(ρL
ρG
)b
; (4.7)
where A is the asymptotic decay rate, x0 the abscissa at which the asymptotic behaviour
begins (virtual origin) and b a power applied to the density ratio. To compare this relation
with a constant density case, the nozzle effective diameter is defined as dn(ρL/ρG )b , used as a
normalisation parameter.
Before presenting the calculation, it should be noted that these relations are constructed for
a gas mixture. Therefore, the mixture velocity u is involved. Here, only the liquid phase is
measured at the centerline, obtained and assimilated from LDV and DTV. However, given the
high density ratio (ρL/ρG = 828), when Y → 1.0 at the centerline, ux,0 → ux,L,0.
Another small difference is about the injection velocity u J in Eq. 4.7. u J is assumed to be a
top-hat flat profile, which is not true in this case. Although this does not change significantly
72
4.1. Experimental observations
the results, Figure 4.3 shows the difference between the bulk velocity u J against the centerline
injection velocity u J ,0 at x/dn = 0 (injection point). These are different because there is a
developed boundary layer inside the injector, and to clarify this point, the results from a k −ϵsimulation case inside the circular injector, along with a calculated power-law 1/7th, are
shown.
-0.6 -0.4 -0.2 0 0.2 0.4 0.6
y (mm)
0
10
20
30
40
50
〈u〉 x
(m/s)
SIM: ux,L k − ǫ
Power-Law: 1/7
Bulk velocity: uJ
Figure 4.3 – Axial velocity profile against radial distance at x/dn = 0.
With these considerations, using b = 0.5, the fitted value for A using Eq. 4.7 is shown in Figure
4.4. The two possible injection velocities are contrasted. Also, only the data in the dispersed
region of the jet is considered, meaning that A is calculated using the data for x/dn > 400, in
the same way as the spreading-rate S before.
0 0.2 0.4 0.6 0.8 1 1.2
x (m)
0.8
1
1.2
1.4
1.6
uJ/u
x,0
LDV Liquid
Fit A = 0.017
(a) Using bulk injection velocity u J =35m/s.
0 0.2 0.4 0.6 0.8 1 1.2
x (m)
0.8
1
1.2
1.4
1.6
uJ,0/u
x,0
LDV Liquid
Fit A = 0.021
(b) Using centerline injection velocity u J ,0 =44m/s.
This yields an average decay rate A = 0.019, nearly 10 times smaller than in the cases reported
by Ruffin et al. [51] and the LDV data from Hussein et al. [30] in an air-air round jet, where
S = 0.094. However, using the same procedure, Stevenin et al. [58] found a similar decay rate
of A = 0.0273, and a spreading rate of S = 0.024.
As no more information is available, the construction of a similarity pattern using the radial
profiles is attempted next. ux,L,0 and y0.5u are used as normalising quantities. Figure 4.5
73
Chapter 4. Results
presents both the axial and lateral liquid velocities, using this procedure.
0 2 4 6 8
y/y0.5u
0
0.2
0.4
0.6
0.8
1
ux,L/u
x,L,0
(a) Axial liquid velocity
0 2 4 6 8
y/y0.5u
-0.01
0
0.01
0.02
0.03
0.04
uy,L/u
x,L,0
(b) Lateral liquid velocity
0 2 4 6 8
y/y0.5u
0
0.2
0.4
0.6
0.8
1
ux,G/u
x,L,0
(c) Axial gas velocity
0 2 4 6 8
y/y0.5u
-0.01
0
0.01
0.02
0.03
0.04
uy,G/u
x,L,0
(d) Lateral gas velocity
x/dn = 400
x/dn = 500
x/dn = 600
x/dn = 700
x/dn = 800
Figure 4.5 – Velocity field from the Liquid and Gas LDV campaign. Profiles against radialdistance from x/dn = 400 to x/dn = 800.
A similarity is found on both components in the liquid part, but the shape differs from a
single phase jet. While the axial component decays slower against the radial distance, the
lateral component remains always positive. This is logic if the liquid velocity follows always
the liquid spreading from the central part of the jet, pointing outwards. Therefore, the slip-
velocity between the phases should always be positive. Indeed, as Figure 4.5 shows, both gas
components fall below the liquid velocity. The entrainment part of the jet is carried out by the
gas phase, but at a much slower intensity than in a single phase jet. And, as expected, there is
no similarity this time on neither of the profiles.
The noise in the profiles, seen as steps far from the centerline, comes from the LDV BP-Filter
setting. As the magnitude of the velocity decreases, the BP-Filter is set to a more narrow span,
this corresponds to discrete cutout frequencies. This effect makes the jumps from one point to
another in the profile, as the configuration is continuously changed to grab the wider possible
band.
To highlight the importance of the average liquid/gas slip-velocity ui ,S , a relation extracted
74
4.1. Experimental observations
from Chapter 2 is repeated here (see Eq. 2.15 on Page 17). From the development of the Favre
averaging process, the Reynolds-averaged phase-velocity fields are related to the turbulent
mass flux in the following expression:
ui ,S = ui ,L − ui ,G =u ′′
i Y ′′
Y (1− Y ); (4.8)
meaning that if a correct estimation of ui ,S is achieved, the form of the turbulent mass fluxu ′′i Y ′′ could be deduced. From Eq. 4.8, ui ,L and ui ,G refer to the Reynolds averaged fields on
the Liquid and Gas phases. However, as discussed before, it is not straightforward to define
the average behaviour of the liquid phase as the average from the LDV or DTV data. Therefore,
the analysis is presented step-by-step.
A better insight on the liquid velocity behaviour can be seen by looking at the DTV profiles.
Despite the lose of spatial precision because of the DOF, as seen before in Chapter 3 (Figure
3.16a on Page 61), the distribution of droplets sizes plays a major role in the reconstruction of
the velocity and the Reynolds stresses fields. The influence of the droplet sizes in the calculated
averaged values is investigated first.
To mimic the previous Liquid-LDV results, the same profiles are constructed using the method
described in Figure 3.20 (Page 67). A first analysis on the influence of the droplet sizes is done
by reconstructing the velocity field using Eq. 4.3 (Average) and Eq. 4.4 (d-Average).
-4 -2 0 2 4
y/y0.5u
0
0.2
0.4
0.6
0.8
1
ux/u
x,0
Average
-4 -2 0 2 4
y/y0.5u
-0.04
-0.02
0
0.02
0.04
uy/u
x,0
Average
x/dn = 400
x/dn = 500
x/dn = 600
x/dn = 700
x/dn = 800
-4 -2 0 2 4
y/y0.5u
0
0.2
0.4
0.6
0.8
1
ux/u
x,0
d-Average
-4 -2 0 2 4
y/y0.5u
-0.04
-0.02
0
0.02
0.04
uy/u
x,0
d-Average
Figure 4.6 – Mean axial and lateral velocities against radial distance. DTV radial profiles fromx/dn = 400 to x/dn = 800.
75
Chapter 4. Results
To check the jet symmetry, the profiles are reconstructed from all the horizontal measurement
points. As Figure 4.6 shows, the same similarity on the profiles is achieved, using the two
types of averages. ux,0 and y0.5u are used as normalising parameters, calculated from the
DTV profiles. Nevertheless, comparing with the Liquid-LDV, it can be noted that although the
similar-profiles keep the same shape, the velocity field obtained by the DTV is not the same.
As mentioned before, there is a difference in the behaviour depending on the class of droplets
by diameter. This is investigated by decomposing the averaged velocity field into the classes
defined in Table 3.4 (Page 63). Then, the mean velocity is calculated using an arithmetic
average of Eq. 4.5.
Using this procedure, Figures 4.7 and 4.8 show the velocity against radial distance profiles, on
all the measurement points, in absolute coordinates. The sub-figure analysis on the shadow
images, along with a sufficient number of objects detected by class, allow the reconstruction
of these detailed profiles.
-40 -20 0 20 40
y (mm)
0
10
20
30
40
ux(m
/s)
x/dn = 400
-40 -20 0 20 40
y (mm)
0
10
20
30
40
ux(m
/s)
x/dn = 500
-40 -20 0 20 40
y (mm)
0
10
20
30
40
ux(m
/s)
x/dn = 600
-40 -20 0 20 40
y (mm)
0
10
20
30
40
ux(m
/s)
x/dn = 700
-40 -20 0 20 40
y (mm)
0
10
20
30
40
ux(m
/s)
x/dn = 800
Class: d[30] < 0.100 mm
Class: 0.100 mm < d[30] < 0.250 mm
Class: 0.250 mm < d[30] < 0.500 mm
Class: 0.500 mm < d[30] < 0.750 mm
Class: 0.750 mm < d[30] < 1.000 mm
Class: 1.000 mm < d[30]
Figure 4.7 – Mean axial velocity by droplet’s class diameter against radial distance. DTV Averageradial profiles from x/dn = 400 to x/dn = 800.
76
4.1. Experimental observations
-40 -20 0 20 40
y (mm)
-1
-0.5
0
0.5
1
uy(m
/s)
x/dn = 400
-40 -20 0 20 40
y (mm)
-1
-0.5
0
0.5
1
uy(m
/s)
x/dn = 500
-40 -20 0 20 40
y (mm)
-1
-0.5
0
0.5
1
uy(m
/s)
x/dn = 600
-40 -20 0 20 40
y (mm)
-1
-0.5
0
0.5
1
uy(m
/s)
x/dn = 700
-40 -20 0 20 40
y (mm)
-1
-0.5
0
0.5
1uy(m
/s)
x/dn = 800
Class: d[30] < 0.100 mm
Class: 0.100 mm < d[30] < 0.250 mm
Class: 0.250 mm < d[30] < 0.500 mm
Class: 0.500 mm < d[30] < 0.750 mm
Class: 0.750 mm < d[30] < 1.000 mm
Class: 1.000 mm < d[30]
Figure 4.8 – Mean lateral velocity by droplet’s class diameter against radial distance. DTVAverage radial profiles from x/dn = 400 to x/dn = 800.
While large droplets tend to maintain the jet average bulk velocity (u j = 35m/s) in the
axial direction, the small ones lag significantly behind. The lateral velocity shows a similar
behaviour, where big droplets tend to escape the central part of the jet, twice as fast as for
the smaller class. Despite using a different class partition, the same behaviour can be seen in
Stevenin [57] case.
This effect is already observed by Prevost et al. [49] in a particle laden tube jet. When particles
come within the same gas jet, their response to the average motion is driven by their capacity
to adapt to the gas flow velocity. So, if the longitudinal average gas velocity decreases, it would
be harder for large particles to adapt, and their average velocity will be higher.
An analogy to this case can be made. Here, a heavy poly-dispersed flow comes from the nozzle,
where droplets meet the gas phase. Dragged by the particles, the gas phase should accelerate
until an equilibrium velocity is reached. Small droplets will adapt quicker to this, since they
are subjected to bigger aerodynamic effects as a function of the local slip-velocity (velocity
seen by the droplets) and their relaxation time τR . This effect can be further investigated by
examining the Reynolds stresses, which are shown next.
As previously mentioned, the gas phase velocity obtained by LDV is not accurate in zones
where a large concentration of liquid droplets is present. Moreover, both LDV and DTV might
have biases related to the measurement volume and DOF. Despite these limitations, the LDV
data allows to estimate ui ,L , ui ,G and ui ,S , where the results show a clear mean slip velocity.
77
Chapter 4. Results
From the DTV side, the incertitude introduced by the DOF is tackled with the analysis by class
of diameter. This allows to have a clear picture on the average behaviour at different scales,
and although it is not explicitly shown here, the relaxation time τR must play a significant role,
as investigated by Ferrand et al. [21].
4.1.2 Reynolds stresses
The same analysis is performed for the calculation of the Reynolds stresses. Since the main
goal is to compare these experimental results with a mixture RANS model, some precisions
must be set before.
The numerical mixture RANS model is Favre-averaged, meaning that if a representation from
the Reynolds stresses by phase is constructed, the combination yields the following relation:
u ′′i u
′′j = Y u
′i ,Lu
′j ,L + (1− Y )u
′i ,G u
′j ,G +
u ′′i Y ′′u ′′
j Y ′′
Y (1− Y ); (4.9)
where the first two terms are the Reynolds-averaged contributions from the two phases, and
the last part is a slip-related component. Actually, using the original relation for the slip-
velocity, ui ,S (Eq. 4.8) into this Eq. 4.9, all contributions to the Favre-averaged Reynolds
stresses can be expressed from known experimental quantities:
• u ′′i u
′′j : Favre averaged Reynolds stresses (or Ri j );
• u′i ,Lu
′j ,L : Liquid Reynolds stresses (or Ri j ,L);
• u′i ,G u
′j ,G : Gas Reynolds stresses (or Ri j ,G );
• ui ,S u j ,S : Slip Reynolds stresses (or Ri j ,S).
Rewriting Eq. 4.9 using these terms, gives:
Ri j = Y Ri j ,L + (1− Y )Ri j ,G + Y (1− Y )Ri j ,S . (4.10)
The overall expression can not be reconstructed, because no experimental results are available
to estimate Y . Nevertheless, the partial contributions are available from the Liquid and Gas
LDV measurements, and the subsequent slip-velocity.
The averaging procedure for the Reynolds stresses on each phase is detailed using the LDV
data. Using the calculated ui ,L (Eq. 4.1) and ui ,G (Eq. 4.2) as center values, the following
estimators are constructed for Ri j ,L and Ri j ,G :
Ri j ,L = 1
n
n∑k=1
(ui ,k∈l i q − ui ,L
)(u j ,k∈l i q − u j ,L
); (4.11)
78
4.1. Experimental observations
Ri j ,G = 1
n
n∑k=1
(ui ,k∈g as − ui ,G
)(u j ,k∈g as − u j ,G
). (4.12)
The slip component, Ri j ,S in Eq. 4.10, is reconstructed from the slip-velocity field defined in
Eq. 4.8, which yields:
Ri j ,S = (ui ,L − ui ,G
)(u j ,L − u j ,G
)= ui ,S u j ,S . (4.13)
Using this procedure, Ri j ,L , Ri j ,G and Ri j ,S contributions are shown in Figure 4.9 for R11, R22
and R12 components.
0 1 2 3 4 5
y/y0.5u
0
0.02
0.04
0.06
0.08
0.1
R11/u
2 L,x,0
0 1 2 3 4 5
y/y0.5u
0
0.002
0.004
0.006
0.008
0.01
R22/u
2 L,x,0
0 1 2 3 4 5
y/y0.5u
0
0.002
0.004
0.006
0.008
0.01
R12/u
2 L,x,0
Rij,L
Rij,G
Rij,S
x/dn = 400
0 1 2 3 4 5
y/y0.5u
0
0.02
0.04
0.06
0.08
0.1
R11/u
2 L,x,0
0 1 2 3 4 5
y/y0.5u
0
0.002
0.004
0.006
0.008
0.01
R22/u
2 L,x,0
0 1 2 3 4 5
y/y0.5u
0
0.002
0.004
0.006
0.008
0.01
R12/u
2 L,x,0
Rij,L
Rij,G
Rij,S
x/dn = 800
Figure 4.9 – Reynolds stresses against radial distance. LDV liquid, gas and slip componentsradial profiles at x/dn = 400 and x/dn = 800.
ux,L,0 and y0.5u are used as normalising parameters. The resulting stresses are presented at
x/dn = 400 and x/dn = 800, to see the evolution from the beginning of the dispersed zone up
to the last profile acquired.
The R11,L component shows a similar behaviour like in a single-phase round jets, reaching its
maximum of 0.08 in the dispersed region [30]. The main difference is in the radial component
R22,L , where a huge anisotropy is found, being R11,L ∼ 15 R22,L . The shear component R12,L
is also small, it reaches R12/k ≈ 1/3 at the end of the liquid velocity profile, where no more
droplets are present. This behaviour differs from the gas-gas variable-density case of Amielh
et al. [2] or the gas-gas constant-density of Hussein et al. [30], where the comparison is shown
79
Chapter 4. Results
in Figure 4.10.
0 2 4 6 8
y/y0.5u
0
0.1
0.2
0.3
0.4
R12,L/k
L x/dn = 400
x/dn = 500
x/dn = 600
x/dn = 700
x/dn = 800
Hussein LDV
0 2 4 6 8
y/y0.5u
0
0.1
0.2
0.3
0.4
R12,G/k
G
Figure 4.10 – Shear stress over turbulent kinetic energy. LDV liquid and gas components.Radial profiles from x/dn = 400 to x/dn = 800.
Here, kL and kG are calculated assuming a cylindrical axisymmetry, meaning:
kL = 1
2
(R11,L +2R22,L
); (4.14)
kG = 1
2
(R11,G +2R22,G
). (4.15)
Since the shear (R12) and lateral (R22) components are the main dissipation terms in momen-
tum equation, these low values could explain the low decay-rate in the centerline velocity and
low spreading-rate. Moreover, it is important to notice that the calculated slip-component
Ri j ,S has the same order of magnitude as the liquid and gas parts (Ri j ,L and Ri j ,G ), meaning
that the complete reconstruction based on these three contributions is important to perform
a proper comparison with the mixture Ri j model.
To characterise the anisotropy of the Reynold stresses, the anisotropy factor is introduced.
Since this liquid round jet presents an axisymmetric behaviour, where R11 ≫ R22 = R33, the
anisotropy factor ⟨A⟩R is defined as:
⟨A⟩R = ⟨R⟩22
⟨R⟩11. (4.16)
Therefore, a low anisotropy factor means a high anisotropy AR ≪ 1.0, and a value close to
AR = 1.0 means an isotropic behaviour.
From the observations made by Stevenin [57], if the Stokes number (St) calculated for the
liquid droplets is small enough, a high anisotropy could be a consequence of the sharp gas
boundary layer created around the poorly atomised liquid jet. However, by looking at the gas
phase Reynolds stresses here, it seems to be the other way around. Indeed, the gas phase data
80
4.1. Experimental observations
presents a lower anisotropy than the liquid phase, meaning that it is the liquid phase which
generates this behaviour.
To investigate further on the source, the DTV data is used. Since the strong anisotropy and
low shear components seem to be maintained throughout the whole dispersed domain, an
analysis at x/dn = 800 is presented. The joint pdf s between the velocity components and the
droplet’s diameters are shown in Figure 4.11 at several radial distances from the centerline,
where using the sub-image partition, the pdf s are constructed by picking the central slices on
each value.
To compare using the whole database, Liquid and Gas LDV are shown on nearly the same
measurement points. The first row corresponds to the LDV Gas (not available at the centerline);
the second row is the LDV Liquid; and the rest are from DTV, decomposed as the pdf s by
velocity component as a function of the droplet equivalent diameter (d[30]). It is clear that
the local dispersion by class of diameters behaves in a very different way on both axes, this is
confirmed by looking at the same pdf shape from the LDV campaign.
The change in the fluctuating behaviour for both velocity components can be seen at the same
time in the Liquid and Gas phases. As the distance from the centerline increases, a tendency
to a more isotropic behaviour can be seen in both the liquid and the gas. By looking at the
decomposition by droplet diameter, it seems that the presence of big droplets close to the
centerline generates a long spectrum of variation for the axial velocities, whereas a less intense
effect is seen in the lateral component.
From this, the Reynolds stresses are obtained using the same averaging procedure used for
the mean velocity estimation: thus, a simple average and a diameter-weighted average. Ri j
flagged as Average, using ui as a centre value from Eq. 4.3, is simply:
Ri j = 1
n
n∑l=1
(ui ,l − ui
)(u j ,l − u j
); (4.17)
and d-Average is calculated as a weighted average by the droplet d[30] diameter, using ui from
Eq. 4.4, meaning:
Ri j =∑n
l=1 d[30],l(ui ,l − ui
)(u j ,l − u j
)∑nl=1 d[30],l
. (4.18)
The results are shown in Figure 4.12, where the same similitude representation appears
using the calculated ux,0 and y0.5u as normalisation parameters, from the corresponding
velocity fields. Once more, no matter the type of average, the profiles fit well a similar relation.
Moreover, the profiles show the same behaviour as those previously obtained by LDV. However,
as expected, the diameter-weighted average produces an impact on the results. As the previous
pdf s show in Figure 4.11, large droplets present little agitation compared to the smaller
ones. This effect is studied by reconstructing the Reynolds stresses by class of diameter. The
averaging procedure is a simple arithmetic average inside the class (k), meaning that the
81
Chapter 4. Results
0
10
20
30
40
ux(m
/s)
-10 -5 0 5 10
uy (m/s)
LDV-L y = 0.0 mm
0
10
20
30
40
ux(m
/s)
-10 -5 0 5 10
uy (m/s)
LDV-L y = 12.0 mm
0
10
20
30
40ux(m
/s)
-10 -5 0 5 10
uy (m/s)
LDV-G y = 12.0 mm
0
10
20
30
40ux(m
/s)
-10 -5 0 5 10
uy (m/s)
LDV-L y = 20.4 mm
0
10
20
30
40
ux(m
/s)
-10 -5 0 5 10
uy (m/s)
LDV-G y = 20.4 mm
0
10
20
30
40
ux(m
/s)
-10 -5 0 5 10
uy (m/s)
LDV-L y = 30.0 mm
0
10
20
30
40
ux(m
/s)
-10 -5 0 5 10
uy (m/s)
LDV-G y = 30.0 mm
0
10
20
30
40
ux(m
/s)
-10 -5 0 5 10
uy (m/s)
DTV y = 0.0 mm
0
10
20
30
40
ux(m
/s)
-10 -5 0 5 10
uy (m/s)
DTV y = 12.6 mm
0
10
20
30
40
ux(m
/s)
-10 -5 0 5 10
uy (m/s)
DTV y = 21.0 mm
0
10
20
30
40
ux(m
/s)
-10 -5 0 5 10
uy (m/s)
DTV y = 30.6 mm
0
10
20
30
40
ux(m
/s)
0.5 1 1.5 2
d30 (mm)
0
10
20
30
40
ux(m
/s)
0.5 1 1.5 2
d30 (mm)
0
10
20
30
40
ux(m
/s)
0.5 1 1.5 2
d30 (mm)
0
10
20
30
40
ux(m
/s)
0.5 1 1.5 2
d30 (mm)
-10
-5
0
5
10
uy(m
/s)
0.5 1 1.5 2
d30 (mm)
-10
-5
0
5
10
uy(m
/s)
0.5 1 1.5 2
d30 (mm)
-10
-5
0
5
10
uy(m
/s)
0.5 1 1.5 2
d30 (mm)
-10
-5
0
5
10
uy(m
/s)
0.5 1 1.5 2
d30 (mm)
Figure 4.11 – Bi-variate histograms normalised as a pdf from LDV and DTV, at x/dn = 800 for:ux -uy , ux -d[30] and uy -d[30].
82
4.1. Experimental observations
-4 -2 0 2 4
y/y0.5u
0
0.02
0.04
0.06
0.08
0.1
R11/u
2 x,0
-4 -2 0 2 4
y/y0.5u
0
1
2
3
4
5
R22/u
2 x,0
×10-3
-4 -2 0 2 4
y/y0.5u
-0.01
-0.005
0
0.005
0.01
R12/u
2 x,0
x/dn = 400
x/dn = 500
x/dn = 600
x/dn = 700
x/dn = 800
Average
-4 -2 0 2 4
y/y0.5u
0
0.02
0.04
0.06
0.08
0.1
R11/u
2 x,0
-4 -2 0 2 4
y/y0.5u
0
1
2
3
4
5
R22/u
2 x,0
×10-3
-4 -2 0 2 4
y/y0.5u
-0.01
-0.005
0
0.005
0.01
R12/u
2 x,0
x/dn = 400
x/dn = 500
x/dn = 600
x/dn = 700
x/dn = 800
d-Average
Figure 4.12 – Reynolds stresses against radial distance. DTV Average and d-Average radialprofiles from x/dn = 400 to x/dn = 800.
83
Chapter 4. Results
values are centred using ui ,(k) (Eq. 4.5):
Ri j ,(k) =1
n
n∑l=1
(ui ,l∈(k) − ui ,(k)
)(u j ,l∈(k) − u j ,(k)
); (4.19)
and shown in Figure 4.13. Since, ux,0 and y0.5u are not defined by class of diameter, the
Reynolds stresses are presented in absolute values, at x/dn = 400 and x/dn = 800.
A clear different behaviour can be observed from the analysis by class of diameter. The shear
and transverse components seem to be more important for very small diameter droplets,
following the gas phase turbulence (see Figure 4.9). This is consistent with the observations
made by Ferrand et al. [21] in a particle-laden jet, who explains this behaviour using the
calculated Stokes number (St ) by class of diameter.
Finally, using these values for the reconstructed Ri j , the anisotropy factor R22,(k)/R11,(k) by
class is constructed and presented in Figure 4.14.
-40 -20 0 20 40
y (mm)
0
0.1
0.2
0.3
0.4
R22/R
11
x/dn = 400
-40 -20 0 20 40
y (mm)
0
0.1
0.2
0.3
0.4
R22/R
11
x/dn = 500
-40 -20 0 20 40
y (mm)
0
0.1
0.2
0.3
0.4
R22/R
11
x/dn = 600
-40 -20 0 20 40
y (mm)
0
0.1
0.2
0.3
0.4
R22/R
11
x/dn = 700
-40 -20 0 20 40
y (mm)
0
0.1
0.2
0.3
0.4
R22/R
11
x/dn = 800
Class: d[30] < 0.100 mm
Class: 0.100 mm < d[30] < 0.250 mm
Class: 0.250 mm < d[30] < 0.500 mm
Class: 0.500 mm < d[30] < 0.750 mm
Class: 0.750 mm < d[30] < 1.000 mm
Class: 1.000 mm < d[30]
Figure 4.14 – Reynolds stresses anisotropy factor (R22/R11) against radial distance. DTV radialprofiles by droplets’ class diameters from x/dn = 400 to x/dn = 800.
These results clearly differs from the gas-gas jet of Hussein et al. [30] or the particle-laden jet
of Ferrand et al. [21], where a high anisotropy is found on the bigger class of particles (dp =80−90µm). Here, the anisotropy seems to reach its maximum in the 100µm < d[30] ≤ 250µm
and 250µm < d[30] ≤ 500µm class, at the more fragmented part of the jet. In contrast, very
84
4.1. Experimental observations
-40 -20 0 20 40
y (mm)
0
20
40
60
R11(m
2/s
2)
-40 -20 0 20 40
y (mm)
0
2
4
6
R22(m
2/s
2)
-40 -20 0 20 40
y (mm)
-6
-4
-2
0
2
4
6
R12(m
2/s
2)
d[30] < 0.100 mm0.100 mm < d[30] < 0.250 mm0.250 mm < d[30] < 0.500 mm0.500 mm < d[30] < 0.750 mm0.750 mm < d[30] < 1.000 mm1.000 mm < d[30]
At x/dn = 400 by Class:
-40 -20 0 20 40
y (mm)
0
20
40
60
R11(m
2/s
2)
-40 -20 0 20 40
y (mm)
0
2
4
6
R22(m
2/s
2)
-40 -20 0 20 40
y (mm)
-6
-4
-2
0
2
4
6
R12(m
2/s
2)
d[30] < 0.100 mm0.100 mm < d[30] < 0.250 mm0.250 mm < d[30] < 0.500 mm0.500 mm < d[30] < 0.750 mm0.750 mm < d[30] < 1.000 mm1.000 mm < d[30]
At x/dn = 800 by Class:
Figure 4.13 – Reynolds stresses against radial distance. DTV radial profiles by droplets’ classdiameters for x/dn = 400 and x/dn = 800.
85
Chapter 4. Results
large and small droplets are more isotropic. As enunciated before, very small droplets could
be more prone to be affected by the gas phase turbulence, therefore they will tend to isotropy
at the external zones. This is also implicitly shown in Figure 4.11, where the combined effect
can be seen by reconstructing the pdf s at several distances from the centerline.
As mentioned before, an estimation of the Stokes Number (St ) could give a better insight on
this behaviour. From the acquired data, the following formulation by class of diameter (k)
could be used:
St(k) =τR
τt=
ρL d 2[30],(k)
18µG
(1+0.15Re0.687
d ,(k)
)y0.5upR11,G
; (4.20)
τt is estimated using the length scale y0.5u and axial standard deviation for the velocity
fluctuations in the gas phase√
R11,G . Red ,(k) is the Reynolds number seen by the droplets of
the class (k):
Red ,(k) =∥ui ,(k) − ui ,G∥d[30],(k)
νG. (4.21)
Here, to represent the velocity seen by the droplets, the mean gas velocity ui ,G is extracted
from the LDV data, and ui ,(k) is the mean velocity of the class (k) from the DTV, with the
corresponding mean diameter d[30],(k). This formulation differs from the one in Eq. 2.41 (page
25) in the sense that it confronts directly the mean slip velocity by class of diameters against
the gas velocity field, whereas the modelled quantity needs the drift part ui ,D (Eq. 2.42) to
account for the slip velocity seen by the droplets.
Because the mean and fluctuating data is not available at the centerline of the jet, a special class
of d[30],(k) ≤ 50µm is created to represent√
R11,G . This is, however, a very strong hypothesis,
because as Figure 4.15 shows at x/dn = 800, despite the similitude far from the centerline of
these two quantities, closer to the axis there is no evidence that small droplets (d[30],(k) ≤ 50µm)
follow the gas-phase fluctuations.
0 10 20 30 40
0
20
40
60
Figure 4.15 – Reynolds stresses against radial distance. DTV and LDV radial profiles at x/dn =800.
86
4.1. Experimental observations
Using these results, the Stokes number St from Eq. 4.20 is calculated along the centerline of
the jet in Figure 4.16. It shows how the particles should react to this gas fluctuating field R11,G .
Here, if St ≫ 1, the droplets are unresponsive to the gas phase fluctuations, and if St ≪ 1, they
should follow the gas phase as tracers.
0 0.2 0.4 0.6 0.8 110
0
101
102
103
Figure 4.16 – Stokes number at the jet centerline by droplets’ class of diameters from x/dn = 400to x/dn = 800.
Despite this strong hypothesis, a clear difference on the behaviour by class of diameter can
be seen. Although the small droplets never reach St ≪ 1, and therefore, they should not be
considered as gas tracers, they are order of magnitudes more responsive to the gas phase
fluctuations than the bigger ones.
This effect could explain the strong anisotropy factor found in the medium sized droplets (see
Figure 4.14). Large droplets are the least influenced by the gas phase fluctuations, they tend
to keep a velocity close to the injection bulk velocity u J = 35m/s, with a turbulent intensity
inherited from the pure liquid-phase (see Figure 4.13). Small droplets follow the gas-phase
fluctuations easily far from the centerline, but they are trapped by the large slip-velocity
induced by large droplets close to the axis. However, medium sizes droplets (100 ≤ d[30],(k) ≤500µm), can have both a wide band of turbulent intensity and be less influenced by the gas
phase fluctuations far from the centerline, boosting the anisotropy on the whole profile as
seen in Figure 4.14.
87
Chapter 4. Results
4.2 Numerical model analysis
The numerical model analysis follows the previous cases definition in Chapter 2. The main
objective at first is to test the behaviour of several variations of the RANS turbulence models.
Later, based on the experimental observations, some specific cases are presented to account
for these observations.
The detailed description of the modelled equations is given in Chapter 2.2. However, to better
illustrate the analysis, some equations are rewritten here along with the study case definition.
A quick description of the simulation case is shown in Figure 4.17. This represents a schematic
view of a longitudinal 2D slice (the real case is 3D).
Nozzle wall
Atmospheric Pressure
Inlet:
x
y
z(Wall Thickness)
Figure 4.17 – Schematic representation of the 3D mesh, including boundary conditions.
From this, the boundary conditions and case set-up are:
• The nozzle diameter is dn = 1.2mm, of length Ln = 50dn and pointing downwards,
aligned with gravity.
• Only water is injected through the nozzle, meaning that u0 = uL,J = 35m/s and Y0 =Y 0 = 1. The air is considered still.
• Turbulence boundary conditions are specified as if there is an infinite, similar, pipe flow
upstream, with a turbulence intensity It = 4%. This yields a k0 = 3.3m2/s2 and ϵ0 =11700m2/s3. The Reynolds stresses are considered isotropic, so Ri j ,0 = 2/3k0δi j m2/s2.
• The simulation time is from t0 = 0 s to t f = 0.3 s. This ensures a full coverage of the
domain, even in the external regions of the jet. The time-step of the simulation is
variable, calculated from the worst case as a function of the local Courant number Co.
To avoid any divergence of the simulation, Co = 0.8 is set as the maximum possible
value.
88
4.2. Numerical model analysis
4.2.1 RANS turbulence model
First, a variation of the RANS turbulence model is analysed. The three previously described
turbulence models are used: k −ϵ (k-Epsilon), Ri j −ϵ (RSM) and Ri j −ϵi j (RSM Variation). To
isolate the behaviour only as a function of the turbulence modelling, the same basic turbulent
mass flux model is used, Ymod0. All of this corresponds to the cases: Case 112, Case 212 and
Case 312 (Page 37).
A first comparison with the experimental results is shown in Figure 4.18. The centerline
evolution of the axial mixture velocity (ux,0) is presented along the centerline liquid velocity
(ux,L,0) from the LDV (Figure 4.18-(a)). Evidently, despite that the comparison of mixture
against liquid velocities might be inconsistent, at the centerline Y ≈ 1.0, which makes ux,0 ≈ux,L,0. Based on the lateral profiles, the calculated half-width of the velocity profiles is
0 0.2 0.4 0.6 0.8 1 1.2
x (m)
0
10
20
30
40
50
〈u〉 x
,0(m
/s)
(a) Axial velocity centerline
ui: k − ǫ
ui: Rij − ǫ
ui: Rij − ǫij
ui,L: LDV
0 0.2 0.4 0.6 0.8 1 1.2
x (m)
0
10
20
30
40
50
y 0.5u(m
m)
(b) Velocity half-width
x/dn > 400 : S = 0.047x/dn > 400 : S = 0.030x/dn > 400 : S = 0.018x/dn > 400 : S = 0.020
0 0.2 0.4 0.6 0.8 1 1.2
x (m)
0
0.2
0.4
0.6
0.8
1
Y
(c) Liquid volume fraction centerline
Y = 0.5 ⇒ Lc/dn = 198
Y = 0.5 ⇒ Lc/dn = 195
Y = 0.5 ⇒ Lc/dn = 198Shadow : Lc/dn = 219
0 0.2 0.4 0.6 0.8 1 1.2
x (m)
0
0.05
0.1
0.15
0.2
√〈R
〉 11,0/〈u〉 x
,0
(d) Axial turbulent intensity centerline
Figure 4.18 – Turbulence models’ benchmark compared to the Liquid LDV.
compared (y0.5u) with experimental reconstruction from the LDV ux,L profiles (Figure 4.18-
(b)). As discussed later, the half-width is defined for the mixture velocity, so a combination
between ux,L and ux,G should be used instead. However, as Y is not directly measured, ux
remains unknown.
The centerline liquid volume fraction is compared to the estimated average liquid column
breakup length, where Y ≈ 0.5 (Figure 4.18-(c)). This is a rough estimation of the behaviour of
the solution, because this hypothesis to separate the dense zone from the dispersed part of
the spray is not always well defined.
89
Chapter 4. Results
Finally, the comparison for the axial turbulent intensity√⟨R⟩11,0/⟨u⟩x,0 is shown in Figure
4.18-(d). Contrary to the mean values, the choice of parameters on the LDV set-up seems to
have a big impact on the calculated fluctuating quantities. Indeed, at x/d = 400, a change
in the PM sensitivity is made, resulting in a jump of the calculated turbulent intensity. To
avoid damaging the PMs, from x/d = 0 to x/d = 400 the gain can not be set to a higher value,
resulting in a biased fluctuating field, as the system is unable to capture small droplets events.
The first observation that can be made is about the momentum diffusion comparison between
the three model approaches. As expected, the introduction of a RANS model that takes into
account some anisotropy on the Reynolds stresses plays a significant role in the overall results.
As Figure 4.18 shows, the Ri j − ϵi j simulation case presents a small decay-rate of the axial
centerline velocity ux,0, bringing the results closer to the experimental points. The explanation
on the mechanism of these results is explained later using the Reynolds stresses fields.
0
0.1
-0.1
0 0.2 0.4 0.6 0.8 1.0 1.2
10.110-210-310-410-5
0
0.1
-0.1
0 0.2 0.4 0.6 0.8 1.0 1.2
10.110-210-310-410-5
0
0.1
-0.1
0 0.2 0.4 0.6 0.8 1.0 1.2
10.110-210-310-410-5
Figure 4.19 – Liquid mass fraction Y field in a mid-plane (z = 0) cutout. Solution from t = 0.1 sto t = 0.3 s.
Because of the U-RANS formulation, and since the solution is taken at a given time, t = 0.3 s,
it can be seen in Figure 4.18-(c) that there is a residual unsteadiness of the solution. Indeed,
90
4.2. Numerical model analysis
Y fluctuates as function of time (therefore, Y does too). This is not due to some lack of
convergence, because the solver converges at every time-step by definition. In fact, these
fluctuations may be a product of the actual jet flapping, like in the work made by Delon [13].
Nevertheless, these fluctuations are not considered to be high enough to affect the analysis
made here, and the unsteadiness is not taken into account in this analysis. To illustrate this, a
complete solution in a mid-plane cutout of the simulated domain is shown in Figure 4.19 for
the Ri j −ϵi j case.
Before going into the comparative analysis, a precision on the use of the turbulent mass flux
model needs to be made. This argument is carried out throughout the whole analysis. Indeed,
Ymod0 reads (from Eq. 2.37):
−ρ u ′′i Y ′′ = µt
σY
∂Y
∂xi. (4.22)
From the model used by Belhadef et al. [5], and the later analysis made by Stevenin et al. [59],
the use of of a turbulent Schmidt number of σY = 5.5 is used and justified. Not only because
it fits well the experimental results on a similar round jet, but because it emulates a strong
anisotropy factor in the principal Reynolds stresses, starting from a more general formulation,
meaning that if a boundary layer approximation is made, then:
−ρ u ′′i Y ′′ ≈ ρu ′′
y Y ′′ =CY ρk
ϵu
′′y
2 ∂Y
∂y; (4.23)
and assuming that u′′y
2 = 0.082k, then using the standard value of CY = 0.1, Ymod1 becomes
Ymod0 with σY = 5.5.
If a RSM case would produce a strong anisotropy, like in the one observed in Figure 4.14,
this artificial σY = 5.5 would not be necessary. However, this is not the case for the two RSM
formulations used here. The anisotropy factor reaches nearly R22/R11 = 0.6 and together with
a high enough shear component R12 ∼ k/3, Ymod1 becomes Ymod0.
Good enough RSM and turbulent mass flux formulations would produce both a strong
anisotropy factor and a low shear component. This would produce at the same time a low
decay rate of the centerline axial velocity and a low spreading rate, as experimentally found
(see Figures 4.1, 4.9 and 4.14).
Having set the same turbulent mass flux model, as previously shown, how the Reynolds
stresses are calculated seems to have a huge effect on the velocity field. This is explicitly shown
in Figure 4.20, where the Reynolds stresses are compared in the dispersed zone of the jet.
91
Chapter 4. Results
-4 -2 0 2 4
y/y0.5u
0
0.02
0.04
0.06
0.08
0.1
〈R〉 1
1/〈u〉2 x
,0
-4 -2 0 2 4
y/y0.5u
0
0.01
0.02
0.03
0.04
0.05
〈R〉 2
2/〈u〉2 x
,0
-4 -2 0 2 4
y/y0.5u
-0.02
-0.01
0
0.01
0.02
〈R〉 1
2/〈u〉2 x
,0
k− ǫ Rij
Rij − ǫ Rij
Rij − ǫij Rij
LDA Rij,L
LDA Rij,G
DTV Rij average
DTV Rij d-average
x/dn = 800
Figure 4.20 – Comparison of the Reynolds stresses against radial distance as a function of theturbulence model at x/dn = 800. Experimental LDV (liquid and gas) and DTV radial profilesare shown as a benchmark.
Indeed, while both the simulation and the experimental case produce nearly the same
turbulent kinetic energy, which for a cylindrical axisymmetry is:
⟨k⟩ = 1
2(⟨R⟩11 +⟨R⟩22 +⟨R⟩33) ≈ 1
2(⟨R⟩11 +2⟨R⟩22) ; (4.24)
the distribution between the principal components is completely different. Indeed, ⟨R⟩11
component has a similar value compared to Hussein et al. [30] or Amielh et al. [2] cases, while
⟨R⟩22 simulation results does not seem to approach the very low experimental values.
Despite all this, since ⟨R⟩12 is indeed lower in the Ri j −ϵi j case, this results in a low momentum
transfer from the axial to the radial direction, decreasing the decay rate of the axial centerline
velocity. A good numerical result would be to reduce considerably ⟨R⟩22 and ⟨R⟩12, while
keeping a high enough ⟨R⟩11.
Although these comparisons are made between the Liquid/Gas Reynolds-averaged values,
against the Favre-averaged model, the orders of magnitude and the relation given by Eq. 4.10
point in a clear direction.
92
4.2. Numerical model analysis
4.2.2 Epsilon equation behaviour
Before examining further the last point, a brief analysis on the use of the Epsilon equation is
developed. The three RANS models used here rely on the same modelled Epsilon equation
(Eq. 2.22) to obtain the turbulent kinetic energy dissipation rate ϵ.
Many versions of this modelled equation exist in different applications of RANS turbulence. A
particular variation to the original version by Jones and Launder [34] is introduced by Pope
[46]. The original version gives good results in a planar-jet configuration, but if the same
parameters are used in a round-jet, the spreading rate S is overestimated. This is called the
round-jet/planar-jet anomaly, it is related to the vortex stretching in the angular direction of
a round-jet. Pope [46] proposes to add an extra source term to account for this, resulting in
good agreement with experimental results.
Dally et al. [9] proposes to use the original equation proposed by Jones and Launder [34], but
with Cϵ1 = 1.60 instead of the original Cϵ1 = 1.44 value. The overall increase of the production
term would produce a similar effect to correct the spreading rate in a round-jet. The analysis
on the application of this modification to a circular multiphase jet is also studied by Stevenin
et al. [59], improving the numerical results on the spreading rate prediction.
To test this behaviour, the contributions at the RHS of the Epsilon equation are studied by
modifying the value of Cϵ1. The study cases are based on the Ri j −ϵ model with Ymod0: Case
211 and Case 212 (see page 38). In Figure 4.21 the following budget is shown for the two cases
at x/dn = 400 in a radial profile:
• Production-1: with Cϵ1 = 1.44 or Cϵ1 = 1.60;
−Cϵ1ϵ
kρu ′′
i u′′j
∂ui
∂x j. (4.25)
• Production-2: with Cϵ4 = 1.0;
−Cϵ4ϵ
ku
′′i
∂p
∂xi. (4.26)
• Destruction: with Cϵ2 = 1.92
−Cϵ2ρϵ2
k. (4.27)
• Dilatation: with Cϵ5 = 1.0
−Cϵ5ρϵ∂uk
∂xk. (4.28)
93
Chapter 4. Results
-2 -1 0 1 2
y/y0.5u
-2000
-1000
0
1000
2000
−/ǫ
(s−1)
(a) Rij − ǫ Cǫ1 = 1.44
-2 -1 0 1 2
y/y0.5u
-2000
-1000
0
1000
2000
−/ǫ
(s−1)
(b) Rij − ǫ Cǫ1 = 1.60
Production-1
Production-2
Destruction
Dilatation
x/dn = 400
Figure 4.21 – Epsilon equation budget against radial distance for two cases: (a) Ri j −ϵ withCϵ1 = 1.44 (Standard value); (b) Ri j −ϵ with Cϵ1 = 1.60 (round-jet correction). Radial profiles atx/dn = 400.
The profiles are divided by ρϵ to account for the relative variation. Cell centre values are also
shown to highlight the mesh quality. Indeed, the original production term (Production-1) is
higher using Cϵ1 = 1.60, creating a higher ϵ, lowering the turbulent kinetic energy k in zones
with high shear.
To check if this effect is also important using the Ri j −ϵi j model with Ymod0, cases Case 311 and
Case 312 (see page 38) are also created. The overall influence is shown in Figure 4.22, where
the axial velocity profiles for all four numerical models are compared against the experimental
results.
0 5 10 15 20 25 30 35 40
y (mm)
0
10
20
30
40
〈u〉 x
(m/s),Y
(−)
Rij − ǫ Cǫ1 = 1.44 ux
Rij − ǫ Cǫ1 = 1.60 ux
Rij − ǫij Cǫ1 = 1.44 ux
Rij − ǫij Cǫ1 = 1.60 ux
Rij − ǫij Cǫ1 = 1.60 Y (×40)
LDV ux,L
LDV ux,G
DTV ux average
DTV ux d-average
x/dn = 400
Figure 4.22 – Mean axial velocity against radial distance as a function of Cϵ1 at x/dn = 400.Experimental LDV (liquid and gas) and DTV radial profiles are shown as a benchmark.
As previously shown before, the Ri j −ϵi j model with Ymod0 predicts not only a better centerline
velocity but at the same time a good spreading rate. From Figure 4.18 (page 89) it might seem
that the spreading rate S is underestimated, but looking at the actual radial profiles, ux should
blend from ux,L to ux,G as a function of the radial distance, following the liquid mass fraction Y
radial profile. This last quantity is however not available from the experimental measurements,
94
4.2. Numerical model analysis
but the numerical solution is shown to illustrate the effect.
4.2.3 Turbulent mass transport
The previous two analyses are centred on the use of an RSM case, coupled with a basic
description for the turbulent mass fluxes (Ymod0). Now, the focus is on this last quantity. Using
several numerical study cases, a comparison between the experimental results against the
numerical solution is made, as a function of several formulations for u ′′i Y ′′ .
Both Ymod0 and Ymod1 are based in the same gradient diffusion hypothesis (Eq. 2.37 and 2.38,
on Page 23). However, as experimentally observed, the mean slip-velocity between the liquid
and gas phases (Eq. 4.8) does not agree with this gradient hypothesis formulation. Indeed, if
ux,S ≫ uy,S , and ∂Y∂x ≪ ∂Y
∂y , then u ′′i Y ′′
HH∝ ∂Y∂xi
.
Throughout the whole set of simulation cases, the second order modelled equation for u ′′i Y ′′ is
solved, but not coupled (Eq. 2.39) with the actual mass transport Eq. 2.10. This is done to have
an estimation on how a solution to this equation would behave, without the complications of
a full coupling, which is analysed later.
From the experimental results, it can be seen that generally u ′′Y ′′ ≫ v ′′Y ′′ . Therefore, the
solution for the second order modelling of u ′′i Y ′′ should produce something like this. An
analysis of the source terms at the RHS of Eq. 2.39 gives an insight on how the solution may
react as a function of a subset of modelled parameters. Indeed, if the system is in equilibrium
and the source and sink terms are dominant, then:
u ′′i Y ′′ = Fi ,D−CF 2
CF 4(1−Y )τR
u ′′i u
′′j
∂Y
∂x j−CF 1
CF 4(1−Y )τR
u ′′j Y ′′ ∂ui
∂x j−CF 3
CF 4(1−Y )τR
Y ′′
ρ
∂p
∂xi; (4.29)
where Fi ,D is the modelled drift velocity, simply expressed as Ymod0. From this, an approxi-
mation can be made, where ∂∂y ≫ ∂
∂x and the pressure-gradient term is neglected, by simply
making CF 3 = 0. Then, the axial and lateral components are:
u ′′x Y ′′ =−CF 2
CF 4(1− Y )τR
u ′′x u
′′y∂Y
∂y− CF 1
CF 4(1− Y )τR
u ′′y Y ′′ ∂ux
∂y(4.30)
u ′′y Y ′′ =− νt
σY
∂Y
∂y− CF 2
CF 4(1− Y )τR u
′′y
2 ∂Y
∂y− CF 1
CF 4(1− Y )τR
u ′′y Y ′′ ∂uy
∂y(4.31)
From this approximation, it can be seen that there is a way to make u ′′Y ′′ ≫ v ′′Y ′′ .
But first, a good solution to for the equivalent mean diameter of droplets d[32] is needed, by
solving Eq. 2.45. This would create a big enough τR close to the centerline, making the far most
RHS terms important. And second, a subset of CF i parameters such as: CF 4 and CF 1 ≫CF 2
95
Chapter 4. Results
are needed too.
To tackle the first point, the solution to the calculated d[32] from the ρΩ solution is shown
in Figure 4.23, along with the d[30] obtain from the DTV experimental measurements at
x/dn = 400.
-40 -30 -20 -10 0 10 20 30 40
y (mm)
0
0.1
0.2
0.3
0.4
0.5
d(m
m)
Rij − ǫij Ymod0: d[32]DTV: d[30]
x/dn = 400
Figure 4.23 – Equivalent diameter of droplets population against radial distance. Radial profilesfrom simulations and DTV at x/dn = 400.
Although the solution does not fit very well the experimental measurements, because not
much attention is taken to the parameters in this study case, the solution is not far off physical
values, therefore it can be used.
Next, choosing CF 1 = 4.0, CF 2 = 0.1, CF 3 = 0.0 and CF 4 = 4.0 (Case 312), the budget for the RHS
terms in Eq. 2.39 is shown in Figure 4.24.
-2 0 2
y/y0.5u
-0.02
-0.01
0
0.01
0.02
Fi/(u
2 x,0/y
0.5u)
u′′
xY′′
−Budget
-2 0 2
y/y0.5u
-2
-1
0
1
2
Fi/(u
2 x,0/y
0.5u)
×10-3 u′′
yY′′
−Budget
F1 : Shear
F2 : Mass
F3 : Pressure
F4 : Drag
x/dn = 600
Figure 4.24 – Turbulent mass transport equation contributions budget against radial distance.Axial and radial components at x/dn = 400.
Where the contributions are:
• F1 Shear: CF 1u ′′
j Y ′′ ∂ui∂x j
;
96
4.2. Numerical model analysis
• F2 Mass: −CF 2u ′′
i u′′j∂Y∂x j
;
• F3 Pressure: −CF 31ρY ′′ ∂p
∂xi;
• F4 Drag: CF 41ρF Dr ag ,i .
From this alternative solution to u ′′i Y ′′ , from now on called Fi , the calculated mean-slip velocity
can also be obtained. This is not a solution to the model however, it is the same Eq. 4.8, but
using the Fi solution. To explicitly show this effect, both formulations are contrasted using the
same RSM case. The comparison between these two solutions is shown in Figure 4.25.
0 1 2 3 4
y/y0.5u
0
0.1
0.2
0.3
ux,S/u
x,0
0 1 2 3 4
y/y0.5u
0
0.01
0.02
0.03
uy,S/u
x,0
Actual
Fi Calculated
LDV
x/dn = 600
Figure 4.25 – Mean slip-velocity against radial distance as a function of Ymod at x/dn = 600.Experimental LDV slip-velocity shown as a benchmark.
Despite the good agreement on the mean velocity fields between the numerical and experi-
mental results, the turbulent mass flux u ′′i Y ′′ does not produce an adequate solution. However,
the Fi solution shows some improvement, where at least ux,S > uy,S . This is also true for the
Reynolds stresses u ′′i u
′′j , where the strong anisotropy cannot be reproduced using this RANS
formulation.
Using these results, a question arises. Could a good Fi solution, coupled with the Reynolds
stresses (via Σi j , Eq. 2.26 on Page 21) generate a strong anisotropy. To investigate this, an
analysis on the contributions to the Ri j equations is made. The objective is to identify how
the anisotropy is generated and what would be the role of Σi j in it.
To illustrate this effect, a first study case without any modification is detailed. It is based on
the same Ri j − ϵi j model, with Ymod0: Case 312 (see page 38). From the Ri j equations, the
contributions at the RHS of Eq. 2.23 are rewritten into the final modelled version and detailed
next:
∂ρu ′′i u
′′j
∂t+∂ρul
u ′′i u
′′j
∂xl− ∂
∂xl
⎡⎣Cs ρk
ϵu ′′
l u′′k
∂u ′′i u
′′j
∂xk
⎤⎦= ρPi j + ρΦi j +Σi j − εi j . (4.32)
97
Chapter 4. Results
where, the first production is:
Pi j =−( u ′′
i u′′k
∂u j
∂xk+ u ′′
j u′′k
∂ui
∂xk
); (4.33)
and the second, variable density production:
1
ρΣi j =
(1
ρG− 1
ρL
)[ u ′′i Y ′′ ∂p
∂x j+u ′′
j Y ′′ ∂p
∂xi
]. (4.34)
The modelled pressure-strain correlation is:
Φi j =φ(sl ow)i j +φ(r api d ,P )
i j +φ(r api d ,Σ)i j ; (4.35)
with, the slow return to isotropy (C1 = 1.8):
φ(sl ow)i j =−C1
ϵ
k
(u ′′i u
′′j −
2
3kδi j
); (4.36)
the rapid Pi j -based redistribution (C2 = 0.6):
φ(r api d ,P )i j =−C2
(Pi j − 1
3Pl lδi j
)(4.37)
and, the rapid Σi j -based redistribution (C3 = 0.75):
φ(r api d ,Σ)i j =−C3
1
ρ
(Σi j − 1
3Σl lδi j
). (4.38)
And finally, the modelled dissipation tensor:
1
ρεi j =
u ′′i u
′′j
kϵ. (4.39)
All these contributions are shown in Figure 4.26, for the components R11, R22 and R12, along
with the calculated P i j from the LDV Liquid and Gas campaigns. As the budget shows, the
shear stress production does not seem to be a source of the anisotropy on its own. Moreover,
the modelled Σi j is a source term only in the R22 component, whatever the value in R11. Even
if a good solution for the turbulent mass flux u ′′i Y ′′ were obtained, coupled with the main
pressure gradient in the axial direction, Σ11 would be negligible compared to the lateral part
Σ22.
This analysis shows that a correct Σi j does not boost the anisotropy, it is the redistribution
term that could play a significant role. Indeed, choosing a C3 ≫ 0.75, all the contribution in
the R22 component could be given to R11. Despite that this variation might be a good start
point to boost the anisotropy, this modification is not investigated in this work.
98
4.2. Numerical model analysis
-2 0 2
y/y0.5u
-0.02
-0.01
0
0.01
0.02
−/(u3 x,0/y
0.5u)
x/dn = 400 R11−Budget
Pij
Σij
ǫij
φ(slow)ij
φ(rapid,P )ij
φ(rapid,Σ)ij
−2R12∂u∂y
Liq
−2R12∂u∂y
Gas
-2 0 2
y/y0.5u
-4
-2
0
2
4
−/(u3 x,0/y
0.5u)
×10-3 x/dn = 400 R22−Budget
Pij
Σij
ǫij
φ(slow)ij
φ(rapid,P )ij
φ(rapid,Σ)ij
−2R22∂v∂y
Liq
−2R22∂v∂y
Gas
-2 0 2
y/y0.5u
-0.02
-0.01
0
0.01
0.02
−/(u3 x,0/y
0.5u)
x/dn = 400 R12−Budget
Pij
Σij
ǫij
φ(slow)ij
φ(rapid,P )ij
φ(rapid,Σ)ij
−R12∂v∂y
− R22∂u∂y
Liq
−R12∂v∂y
− R22∂u∂y
Gas
Figure 4.26 – Reynolds stresses equations budget against radial distance at x/dn = 400 forthe Ri j − ϵi j Ymod0 case. Experimental LDV (liquid and gas) radial profiles are shown as abenchmark.
99
Chapter 4. Results
4.2.4 Fully-coupled turbulent model
The full coupled turbulent model comes from the solution of Eq. 2.39 and the incorporation of
this solution back to Eq. 2.12 and the momentum equation solver. All the system of equations
is fully coupled. To do this, a detail on the solution of the second order modelled equation foru ′′i Y ′′ is given first. Later, some precisions are given about the numerical solver scheme, which
are necessary to attain a stable and converged solution at each time-step.
As previously shown, the RHS of Eq. 4.40 is an important source term in the pressure Eq. 2.53
(see page 30).
∂ρY
∂t+ ∂ρui Y
∂xi=−∂ρ
u ′′i Y ′′
∂xi. (4.40)
In the second order model, u ′′i Y ′′ are obtained by solving Eq. 2.39. Then, Eq. 4.40 could
be solved by simply introducing this solution into the RHS. However, this situation is not
numerically stable for an iterative solver, where the solution is calculated starting from the
previous one. Without an implicit term in the diffusion part for Y , it is only up to the numerical
diffusion to maintain the equation in a parabolic form. To overcome this, a blended solution
is proposed between Ymod0 and Ymod2, where the fluxes are simply:
u ′′i Y ′′ = FB
u ′′i Y ′′
Ymod0+ (1−FB ) u ′′
i Y ′′Ymod2
. (4.41)
If this form is introduced into to Eq. 4.40, the final equation to solve is:
∂ρY
∂t+ ∂ρui Y
∂xi= FB
∂
∂xi
(µt
σY
∂Y
∂xi
)− (1−FB )
∂ρFi
∂xi; (4.42)
where Fi is the solution to Eq. 2.39 (Ymod2) on Case 312, and FB = 0.1 is the blend parameter
to set between the two modelled forms.
To compare with the previous model and experimental results, the axial velocity against radial
distance at x/dn = 400 is presented in Figure 4.27. The results show no big improvement from
the velocity field point-of-view. Indeed, the increase on the mean axial velocity is a direct
consequence of the better representation of the axial slip-velocity ux,S . However, as Figure
4.25 showed, the solution for the turbulent mass fluxes Fi produces a negligible slip-velocity
ui ,S , compared to the one obtained by LDV.
100
4.2. Numerical model analysis
0 5 10 15 20 25 30 35 40
y (mm)
0
10
20
30
40
ux(m
/s)
Rij − ǫij YMod0
Rij − ǫij YMod2
LDA uL,x
LDA uG,x
DTV ux average
DTV ux d-average
x/dn = 400
Figure 4.27 – Comparison of the mean axial velocity against radial distance as a function ofYmod at x/dn = 400. Experimental LDV (liquid and gas) and DTV radial profiles are shown as abenchmark.
The set of parameters used in the Ymod2 solution is not carefully investigated. Despite that the
choice made for Case 312 allows to generate a slip-velocity in the axial direction (ux,S), contrary
to Ymod0 or Ymod1, a fine tuning of CF 1, CF 2, CF 3 and CF 4 may produce a better solution.
Finally, to see the overall behaviour of this modelling approach, the solution for centerline
velocity ux,0, spreading (y0.5u) and centerline volume fraction Y 0 are shown in Figure 4.28.
0 0.2 0.4 0.6 0.8 1 1.2
x (m)
0
10
20
30
40
50
ux,0(m
/s)
(a) Axial mixture velocity
k− ǫ YMod0
Rij − ǫ YMod0
Rij − ǫij YMod0
Rij − ǫij YMod2
LDV-Liquid
0 0.2 0.4 0.6 0.8 1 1.2
x (m)
0
10
20
30
40
50
y 0.5u(m
m)
(b) Mixture velocity half-width
x/dn > 400 : S = 0.047x/dn > 400 : S = 0.030x/dn > 400 : S = 0.018x/dn > 400 : S = 0.020x/dn > 400 : S = 0.020
Figure 4.28 – Turbulence mass transport models’ benchmark. (a) Axial velocity along thecenterline; (b) Axial velocity half-width; (c) Liquid volume fraction along the centerline.
101
Chapter 4. Results
The effect produced on the mean spreading-rate and the centerline velocity decay-rate may
not be significant. However, the increase of the slip-velocity is produced by the increase of the
turbulent mass flux, modifying significantly the solution of the liquid volume fraction. This is
not necessarily a bad solution, because the only experimental reference is the hypothesis that
at the breakup point Y = 0.5, which may not be necessarily true in this liquid jet.
To combine and to solve this fully-coupled model approach requires a lot of considerations,
from the modelling of the actual physics and from numerical stability. Originally, this approach
is conducted to try to generate a more case-independent formulation, relying less on modelled
quantities and the choice of parameters. However, these efforts seem not to pay off, as even a
second-order closure model, fully coupled with the momentum solver also needs fine tuning.
102
4.2. Numerical model analysis
Summary
The combined results from the experimental campaigns and the numerical simulations are
presented in this chapter. From this, the following points could summarise the analysis:
• The experimental results obtained from the LDV and DTV campaigns are presented
first. From this analysis, some basic parameters like the axial velocity decay-rate and the
spreading rate are calculated. These values are compared to the behaviour of other cases
from the literature. Although the density ratio of this case study is high (ρL/ρG = 829),
these results seem to be in accordance with other liquid round-jets cases, like in diesel
injectors.
• The reconstruction of the velocity and fluctuation fields is based on the two separate
LDV campaigns, the aim is to capture the liquid phase and gas phase around it. To
complement, the DTV provides a fine decomposition of the liquid fields, by class of
droplet sizes. The results show a non-negligible average slip-velocity ui ,S between the
phases. This quantity plays a significant role in the reconstruction of the Favre-averaged
Reynolds stresses.
• From the reconstruction of liquid and gas Reynolds stresses, a low anisotropy factor
of the principal components is found on both phases, meaning a high anisotropy. In
the liquid part, this value can be as small as R22,L/R11,L ∼ 0.05, whereas in the gas
phase, it can reach R22,G /R11,G ∼ 0.1. These results differ significantly from the ones
found in constant density round jets, where R22/R11 ∼ 0.6. The decomposition of the
fluctuating fields by class of droplets gives a clue on the mechanism that might produce
this behaviour. Indeed, the results show a drastic change in the anisotropy factor: big
droplets seem to keep the same fluctuating energy from the liquid core, but a high
velocity as well; as the jet breaks into droplets of smaller sizes, they seem to be more
and more affected by the slip-velocity between the big droplets and gas phase, creating
a wider band for the fluctuations to operate in the axial direction; finally, the smallest
droplet group (d[30] < 100µm) seems to follow purely the gas phase fluctuations, at
the external zones of the gas entrainment. Some authors explain this behaviour by
calculating the Stokes number St , however, this analysis is not presented here.
• The experimental results serve as a baseline to compare the constructed simulation
cases. The analysis is centred on the behaviour of a RSM turbulence formulation,
nevertheless, a basic k −ϵ model is also shown for comparison purposes. The velocity
field obtained using the RSM turbulence is very close to the experimental observations,
however, this result is obtained assuming a very low anisotropy factor, following the
experimental observations. This transforms into a very high turbulent Schmidt number
of σY = 5.5, which is far from the most common use of σY ≈ 0.9.
• Despite the inclusion of variable density effects into the RSM model, even with the
use of a second-order solution for the turbulent mass fluxes u ′′i Y ′′ , the high anisotropy
experimentally observed can not be reproduced in the simulation cases. An analysis
103
Chapter 4. Results
on the budget inside the Ri j equations shows that the Σi j production only works in the
lateral R22 component. This result points out that maybe it is the redistribution part,
modelled from the pressure-strain correlation, that might play a significant role in the
source of the anisotropy.
104
Summary, conclusions andperspectives
This short chapter is purely dedicated to the general conclusions of this work. Although a
series of partial conclusions are already presented on each chapter, the general view presented
here is made to wrap-up the combined experimental and numerical approaches.
The study of the performance of sprinklers/sprayers in agricultural applications is a contin-
uous development research. New regulations aim to both reduce water consumption on
irrigation applications, and to minimise ambient pollution when crop protection products
are applied to cultures. These research subjects are carried out at IRSTEA Montpellier cen-
tre, where technical, normative, experimental and theoretical approaches are developed in
conjunction with public and private institutions.
The experimental and numerical approaches treated here are only one small part of the
vast research applied to agricultural sprinklers/sprayers at IRSTEA in collaboration with
IRPHE. From these particular activities, the following points are extracted to summarise and
to conclude this work:
• Based on previous observations, and to simplify the experimental conditions, a particu-
lar case scenario is created to study the atomization and dispersion on an agricultural-
like jet, where purified water is injected into stagnant air. From this, a circular nozzle
of diameter dn = 1.2mm and length Lc /dn = 50 is created. The injection average bulk
velocity is set to u J = 35m/s. This geometry, fluid properties and operating conditions
produce a turbulent atomization regime.
• The atomization and dispersion are first investigated using numerical CFD simulations.
Here, the liquid jet is represented as a variable-density single-fluid Favre-averaged
mixture. Since the size of the problem is relatively big, compared to other applications
like fuel injectors, the advantage of such modelling technique is that there is no need
to represent every length scale present on the flow, therefore saving on computational
resources, but at the expense of model completeness.
105
Summary, conclusions and perspectives
• The mixture model is successfully implemented using the OpenFOAM CFD code. It
allows to represent the solution in an arbitrary 3D mesh, running under a custom U-
RANS solver. The flexibility of the programming philosophy behind the code allows to
easily implement several turbulence models: k −ϵ and RSM ; both including variable
density effects from the mixture model. Also, first-order and second-order closures for
the turbulent mass flux are implemented in a fully coupled solver. The quasi-multiphase
approach is tackled by the use of a transport equation for the mean interface area per
unit volume quantity.
• To have a benchmark baseline for the model, an experimental campaign is carried out.
LDV and Shadowgraphy optical techniques are used to measure the mean velocity and
fluctuating fields. LDV is used to capture the liquid field and the gas around it, by seeding
small olive-oil particles as tracers. In a separate campaign, DTV from the shadow images
is applied to the disperse part of the jet, x/dn > 400, adding more information to the
liquid phase related to the distribution of droplet sizes.
• A very specific LDV configuration is used to perform the data acquisition on each
phase. A first measurement campaign is performed only in the liquid, where the results
are assimilated to the velocity and fluctuating fields of the jet’s liquid phase. For the
gas phase, since the liquid droplets might interfere with the olive-oil tracers, some
considerations have to be taken. First, it is found that the Doppler signal detected
from the relatively big water droplets (d > 30µm) is considerably higher than the one
produced by the small olive-oil tracers (d ∼ 1µm), making the threshold of the LDV
burst signal a good candidate to separate the gas from the liquid signal. Although the
available LDV equipment does not allow to perform such separation, a combined narrow
BP-Filter, higher acceptable SNR and higher sensitivity on the PM gain allow to eliminate
most of the droplet events captured along with the tracers. The resulting signal is not
completely depurated from the liquid droplets. However, having the LDV results only
in the liquid phase as a comparison, these are considered to be closer to the expected
behaviour of a gas velocity signal, and not from a mixture of liquid droplets and gas
tracers. This combined technique allows to reconstruct the liquid and gas velocity fields
along with the Reynolds stresses within an acceptable margin of error.
• Shadow images are used to run a custom DTV algorithm developed and implemented
in MATLAB. A shadow strobe system from Dantec Dynamics is used to capture the
projected shadow of the liquid jet/droplets into a high-speed CCD camera. The first
part of the algorithm detects and extracts droplet’s contours from the acquired 12-bit
grey-scale shadow images, even if they are out-of-focus. The second part performs a
matching between the centroids of these contours to estimate the velocity from two
consecutive frames. To account for the out-of-focus droplets, a calibration procedure is
performed on opaque discs of a known diameter. This procedure gives an equivalent
diameter correction for the out-of-focus objects, along with their relative position as
a function of the detected contrast ratio and edge gradient. The results of the DTV
algorithm are the velocity and fluctuating fields, decomposed by the estimated droplets
106
Summary, conclusions and perspectives
diameters.
• The numerical model seems to approach the experimental results when the mean
velocity fields are contrasted. The calculated values for the centerline velocity decay rate
and the spreading rate are in accordance with values found in the literature. However,
no good agreement is found when comparing the fluctuating fields. Indeed, one of the
main motivations of this work is the use of a RSM to take into account the anisotropy on
the Reynolds stresses. However, several variations of this approach, even with a second-
order closure for the turbulent mass flux, do not seem to approach the experimentally
found anisotropy, where R22/R11 ∼ 0.05.
• A close analysis on the source mechanism that might produce the anisotropy of the
Reynolds stresses in the RSM formulation is studied. If this effect is not present in
constant density or slightly variable density flow, the mechanism of production must
be a consequence of the large density ratio of this case (ρL/ρG = 828). By examining
the source terms in the Ri j equations from the RSM, the redistribution part of the
variable density production term Σi j , issued from the modelling of the pressure-strain
correlation, seems to be a good candidate to investigate. The coupling of the mean
pressure gradient and the turbulent mass fluxes, that generates Σi j , seems to act only
on the lateral R22 component, no matter if a first-order or second-order formulation is
used. This means that the redistribution part should be the only option to boost the
anisotropy under this formulation.
Although the comparative analysis from the two approaches does not always give good results,
it is considered that both add a good amount of information to the understanding of this study
case. Nevertheless, there are several topics that could be treated to improve the analysis, such
as:
• The calibration procedure on the shadow images is not applied to the measured droplet
population. As the results show, the average and fluctuating quantities are strongly
dependent on the granulometry. Therefore, a proper distribution of the droplet sizes
must be obtained, by eliminating the biases related to the DOF and sizes estimation.
However, if done so, more images would be needed to reconstruct a set of well converged
average fields, since more droplets are likely to be rejected from the analysis.
• With a good estimation of the droplet population and distribution, a good estimation of
the average liquid mass fraction Y can be made from the shadow images. Stevenin [57]
work shows a good agreement between the data acquired using an OP and the estimation
made by the DTV system, where Y is estimated by placing the volume occupied by the
droplets inside the calculated DOF.
• The data obtained by the LDV on the gas phase is considered to be only an estimation.
Indeed, the contamination of liquid droplets inside the population of gas tracers events
results induces an underestimation of the Liquid-Gas slip-velocity ui ,S . Newer LDV BSA
systems allow to carefully discriminate events by the Doppler burst pedestal intensity.
107
Summary, conclusions and perspectives
The use of such a system could improve the accuracy and precision of the gas results,
relying less on artificial filtering that might introduce several biases.
• Having a good droplet population and a solid estimation of ui ,S , the Stokes number St
can be calculated by droplet class of diameter. This quantity would allow to have a better
explanation on the droplet’s response to turbulent fluctuations as a function of their
sizes. This mechanism seems to be a good candidate to explain the strong anisotropy of
the Reynolds stresses.
• From the analysis of the Ri j equations budget, there is only one possible way to boost
the anisotropy of the Reynolds stresses. The production term associated to the variable
density formulation, Σi j , is only significantly important in the lateral direction, despite
that ux,S ≫ uy,S . Using only a linear pressure-strain correlation model, it is the redis-
tribution part of Σi j which could play a significant role in the anisotropy production.
Although there is no more information to support this, increasing C3 ≫ 0.75 would kill
the source term in the lateral direction, creating an artificial source in the axial one.
Based on these perspectives, to conduct a new study case would require a new measurement
campaign, along with new simulation cases. To carry on these activities simultaneously is very
time consuming, and would also require new experimental equipment and HPC availability.
Finally, as a general conclusion, a great amount of effort is put to carefully implement and to
solve the numerical cases constructed, along with a detailed experimental campaign. These
two activities, carried out simultaneously, allow to see the results from a perspective that gives
a valuable feedback in both directions.
108
Bibliography
[1] M. Al Heidary, J. Douzals, C. Sinfort, and A. Vallet. Influence of spray characteristics
on potential spray drift of field crop sprayers: A literature review. Crop Protection, 63:
120–130, 2014.
[2] M. Amielh, T. Djeridane, F. Anselmet, and L. Fulachier. Velocity near-field of variable
density turbulent jets. International Journal of Heat and Mass Transfer, 39(10):2149–2164,
1996.
[3] A. Andreini, C. Bianchini, S. Puggelli, and F.-X. Demoulin. Development of a turbulent
liquid flux model for Eulerian–Eulerian multiphase flow simulations. International
Journal of Multiphase Flow, 81:88–103, 2016.
[4] P.-A. Beau. Modélisation de l’atomisation d’un jet liquide : application aux sprays Diesel.
Doctoral thesis, Université de Rouen, 2006.
[5] A. Belhadef, A. Vallet, M. Amielh, and F. Anselmet. Pressure-swirl atomization: Modeling
and experimental approaches. International Journal of Multiphase Flow, 39:13–20, 2012.
[6] A. Boutier. Vélocimétrie laser pour la mécanique des fluides. Lavoisier, 2012. ISBN
978-2-7462-8821-8.
[7] A. Cartellier. Optical probes for local void fraction measurements: Characterization of
performance. Review of Scientific Instruments, 61(2):874–886, 1990.
[8] P. Chassaing, R. A. Antonia, F. Anselmet, L. Joly, and S. Sarkar. Variable Density Fluid
Turbulence, volume 69 of Fluid Mechanics and Its Applications. Springer Netherlands,
2002. ISBN 978-90-481-6040-2.
[9] B. B. Dally, D. F. Fletcher, and A. R. Masri. Flow and mixing fields of turbulent bluff-body
jets and flames. Combustion Theory and Modelling, 2(2):193–219, 1998.
[10] G. G. Daves, R. O. Buckius, J. E. Peters, and A. R. Schroeder. Morphology Descriptors
of Irregularly Shaped Particles from Two-Dimensional Images. Aerosol Science and
Technology, 19(2):199–212, 1993.
[11] M. De Luca. Contribution à la modélisation de la pulvérisation d’un liquide phytosanitaire
en vue de réduire les pollutions. Doctoral thesis, Aix-Marseille II, 2007.
109
Bibliography
[12] M. De Luca, A. Vallet, and R. Borghi. Pesticide atomization modeling for hollow-cone
nozzle. Atomization and Sprays, 19(8):741–753, 2009.
[13] A. Delon. Instabilité de flapping: origine et effets sur la structure et le spray d’un jet atomisé.
Doctoral thesis, Université Grenoble Alpes, 2016.
[14] F.-X. Demoulin, P.-A. Beau, G. Blokkeel, A. Mura, and R. Borghi. A new model for turbulent
flows with large density fluctuations: Application to liquid atomization. Atomization and
Sprays, 17(4), 2007.
[15] J. Desantes, J. García-Oliver, J. Pastor, A. Pandal, E. Baldwin, and D. Schmidt. Coupled/de-
coupled spray simulation comparison of the ECN spray a condition with the Σ-Y Eulerian
atomization model. International Journal of Multiphase Flow, 80:89–99, 2016.
[16] C. Dumouchel. On the experimental investigation on primary atomization of liquid
streams. Experiments in Fluids, 45(3):371–422, 2008.
[17] B. Duret, J. Reveillon, T. Menard, and F. F.-X. Demoulin. Improving primary atomization
modeling through DNS of two-phase flows. International Journal of Multiphase Flow, 55:
130–137, 2013.
[18] H. A. El-Asrag and M. Braun. Effect of turbulence non-isotropy modeling on spray
dynamics for an evaporating Acetone spray jet. International Journal of Multiphase Flow,
68:100–120, 2015.
[19] N. Fdida. Développement d’un système de granulométrie par imagerie Application aux
sprays larges et hétérogènes. Ph.d thesis, Université de Rouen, 2008.
[20] N. Fdida and J.-B. Blaisot. Drop size distribution measured by imaging: determination
of the measurement volume by the calibration of the point spread function. Meas. Sci.
Technol., 21(2):025501, 2010.
[21] V. Ferrand, R. Bazile, J. Borée, and G. Charnay. Gas–droplet turbulent velocity correlations
and two-phase interaction in an axisymmetric jet laden with partly responsive droplets.
International Journal of Multiphase Flow, 29(2):195–217, 2003.
[22] J. H. Ferziger and M. Peric. Computational Methods for Fluid Dynamics. Springer-Verlag
Berlin Heidelberg, 2002. ISBN 3540420746.
[23] M. M. Gibson and B. E. Launder. Ground effects on pressure fluctuations in the
atmospheric boundary layer. Journal of Fluid Mechanics, 86(03):491, 1978.
[24] S. Gold, A. Rangarajan, C. Lu, S. Pappu, and E. Mjolsness. New algorithms for 2D and
3D point matching: pose estimation and correspondence. Pattern Recognition, 31(8):
1019–1031, 1998.
[25] C. J. Greenshields. OpenFOAM 2.4.0 Programmer’s Guide. OpenFOAM Foundation Ltd.,